Token \u5b8c\u5168\u79d1\u666e\uff1aLLM \u8c03\u7528\u6d88\u8017\u4e0e\u8ba1\u8d39\u6307\u5357<\/title>\n<style>\n:root {\n --bg: #f8f9fb; --surface: #ffffff; --ink: #1a1a2e; --muted: #5a6578;\n --rule: #e2e5ea; --accent: #2563eb; --accent2: #059669; --accent3: #8b5cf6;\n --code-bg: #1e293b; --code-text: #e2e8f0;\n --warn-bg: #fffbeb; --warn-border: #f59e0b;\n --err-bg: #fef2f2; --err-border: #ef4444;\n --ok-bg: #ecfdf5; --ok-border: #10b981;\n --info-bg: #eff6ff; --info-border: #3b82f6;\n --tip-bg: #f5f3ff; --tip-border: #8b5cf6;\n}\n[data-theme=\"dark\"] {\n --bg: #0f172a; --surface: #1e293b; --ink: #f1f5f9; --muted: #94a3b8;\n --rule: #334155; --accent: #60a5fa; --accent2: #34d399; --accent3: #a78bfa;\n --code-bg: #020617; --code-text: #cbd5e1;\n --warn-bg: #2a1b0f; --warn-border: #f59e0b;\n --err-bg: #2a0f0f; --err-border: #ef4444;\n --ok-bg: #0f2a1b; --ok-border: #10b981;\n --info-bg: #172554; --info-border: #3b82f6;\n --tip-bg: #1a1530; --tip-border: #8b5cf6;\n}\n*{margin:0;padding:0;box-sizing:border-box}\nbody{font-family:-apple-system,BlinkMacSystemFont,\"Segoe UI\",Roboto,\"Helvetica Neue\",\"Noto Sans SC\",\"PingFang SC\",\"Microsoft YaHei\",sans-serif;background:var(--bg);color:var(--ink);line-height:1.75;font-size:15px;transition:background .3s,color .3s}\n.container{max-width:920px;margin:0 auto;padding:5rem 1.25rem 3rem}\n\n\/* Header *\/\nheader{position:fixed;top:0;left:0;right:0;z-index:100;background:var(--surface);border-bottom:1px solid var(--rule);padding:.85rem 0;transition:background .3s}\n.header-inner{max-width:920px;margin:0 auto;padding:0 1.25rem;display:flex;justify-content:space-between;align-items:center}\n.logo{font-size:1.15rem;font-weight:700;color:var(--accent);display:flex;align-items:center;gap:.5rem}\n.logo svg{width:24px;height:24px}\n.controls{display:flex;gap:.6rem;align-items:center}\n.lang-switch{display:flex;background:var(--bg);border:1px solid var(--rule);border-radius:6px;overflow:hidden;transition:background .3s}\n.lang-switch button{border:none;background:transparent;color:var(--muted);padding:.35rem .75rem;cursor:pointer;font-size:.82rem;transition:all .2s}\n.lang-switch button.active{background:var(--accent);color:#fff}\n.theme-toggle{background:var(--bg);border:1px solid var(--rule);border-radius:6px;padding:.35rem .6rem;cursor:pointer;color:var(--muted);font-size:1rem;line-height:1;transition:all .2s}\n\n\/* Page *\/\n.page-header{margin-bottom:2.5rem;text-align:center}\n.page-header h1{font-size:1.8rem;font-weight:800;letter-spacing:-0.02em;margin-bottom:.5rem}\n.page-header p{color:var(--muted);font-size:1rem;max-width:640px;margin:0 auto}\n.badge{display:inline-block;background:var(--accent3);color:#fff;padding:.25rem .75rem;border-radius:999px;font-size:.75rem;font-weight:600;margin-bottom:.8rem;text-transform:uppercase;letter-spacing:.05em}\n\n.card{background:var(--surface);border:1px solid var(--rule);border-radius:12px;padding:1.75rem 2rem;margin-bottom:1.5rem;transition:background .3s}\n.card h2{font-size:1.25rem;font-weight:700;margin-bottom:1rem;padding-bottom:.5rem;border-bottom:2px solid var(--accent);display:inline-block}\n.card h3{font-size:1.05rem;font-weight:600;margin:1.5rem 0 .5rem;color:var(--accent)}\n.card h3:first-child{margin-top:0}\n.card p,.card li{color:var(--muted);margin-bottom:.6rem}\n.card ul,.card ol{padding-left:1.5rem;margin-bottom:.75rem}\nli{margin-bottom:.4rem}\n\npre{background:var(--code-bg);color:var(--code-text);padding:1.1rem 1.25rem;border-radius:8px;overflow-x:auto;font-family:\"SF Mono\",Monaco,\"Cascadia Code\",Consolas,monospace;font-size:.85rem;line-height:1.6;margin:.75rem 0}\ncode{background:var(--bg);color:var(--accent);padding:.15rem .4rem;border-radius:4px;font-family:\"SF Mono\",Monaco,\"Cascadia Code\",Consolas,monospace;font-size:.87em}\npre code{background:transparent;color:inherit;padding:0}\n.cm{color:#94a3b8}\n\n.cmd{display:flex;align-items:center;gap:.75rem;background:var(--code-bg);color:var(--code-text);padding:.8rem 1.2rem;border-radius:8px;font-family:\"SF Mono\",Monaco,\"Cascadia Code\",Consolas,monospace;font-size:.87rem;margin:.6rem 0;overflow-x:auto}\n.cmd .p{color:#6ee7b7;user-select:none;flex-shrink:0}\n.cmd .c{white-space:nowrap}\n\n.callout{padding:1rem 1.25rem;border-radius:8px;margin:1rem 0;border-left:4px solid}\n.callout.warn{background:var(--warn-bg);border-color:var(--warn-border)}\n.callout.err{background:var(--err-bg);border-color:var(--err-border)}\n.callout.ok{background:var(--ok-bg);border-color:var(--ok-border)}\n.callout.info{background:var(--info-bg);border-color:var(--info-border)}\n.callout.tip{background:var(--tip-bg);border-color:var(--tip-border)}\n.callout strong{display:block;margin-bottom:.3rem;color:var(--ink)}\n\n.table-wrap{overflow-x:auto;margin:1rem 0}\ntable{width:100%;border-collapse:collapse;font-size:.89rem;min-width:500px}\nth,td{padding:.6rem .85rem;text-align:left;border-bottom:1px solid var(--rule)}\nth{background:var(--bg);font-weight:600;color:var(--ink);white-space:nowrap}\ntr:hover td{background:var(--bg)}\n\n.tag{display:inline-block;padding:.15rem .5rem;border-radius:4px;font-size:.75rem;font-weight:600;margin-right:.2rem}\n.tag-r{background:var(--accent);color:#fff}\n.tag-g{background:var(--accent2);color:#fff}\n.tag-o{background:var(--warn-bg);color:#d97706;border:1px solid var(--warn-border)}\n\n.compare-grid{display:grid;grid-template-columns:1fr 1fr;gap:1rem;margin:1rem 0}\n.compare-box{border:1px solid var(--rule);border-radius:8px;padding:1.1rem;background:var(--surface)}\n.compare-box .box-title{font-weight:700;font-size:.88rem;margin-bottom:.5rem;color:var(--accent)}\n\n.summary-box{background:var(--surface);border:2px solid var(--accent);border-radius:12px;padding:1.5rem 1.75rem;margin-top:1.5rem}\n.summary-box h2{border-bottom-color:var(--accent2);margin-top:0}\n.s-row{display:flex;gap:1.5rem;margin:.6rem 0;align-items:flex-start}\n.s-label{flex-shrink:0;font-weight:700;font-size:.9rem;color:var(--accent);min-width:120px}\n.s-desc{color:var(--muted);font-size:.92rem}\n\n.lang-section{display:none}\n.lang-section.active{display:block}\n\nfooter{text-align:center;color:var(--muted);font-size:.8rem;margin-top:2rem;padding-top:1rem;border-top:1px solid var(--rule)}\n\n@media(max-width:640px){\n .compare-grid{grid-template-columns:1fr}\n .page-header h1{font-size:1.5rem}\n .card{padding:1.25rem 1.5rem}\n .s-row{flex-direction:column;gap:.3rem}\n .header-inner{flex-wrap:wrap;gap:.5rem}\n}\n<\/style>\n<\/head>\n<body>\n\n<header>\n <div class=\"header-inner\">\n <div class=\"logo\">\n <svg viewBox=\"0 0 24 24\" fill=\"none\" stroke=\"currentColor\" stroke-width=\"2\" stroke-linecap=\"round\" stroke-linejoin=\"round\">\n <circle cx=\"12\" cy=\"12\" r=\"10\"\/><path d=\"M12 6v6l4 2\"\/>\n <\/svg>\n Token Guide\n <\/div>\n <div class=\"controls\">\n <div class=\"lang-switch\">\n <button class=\"active\" onclick=\"setLang('zh')\" id=\"btn-zh\">\u4e2d\u6587<\/button>\n <button onclick=\"setLang('en')\" id=\"btn-en\">English<\/button>\n <\/div>\n <button class=\"theme-toggle\" onclick=\"toggleTheme()\" title=\"Toggle theme\">☼<\/button>\n <\/div>\n <\/div>\n<\/header>\n\n<div class=\"container\">\n\n\n<div class=\"lang-section active\" id=\"lang-zh\">\n\n<div class=\"page-header\">\n <span class=\"badge\">Token \u79d1\u666e<\/span>\n <h1>Token \u5b8c\u5168\u79d1\u666e\uff1aLLM \u8c03\u7528\u7684\u6d88\u8017\u4e0e\u8ba1\u8d39<\/h1>\n <p>\u4ece\u96f6\u7406\u89e3 Token \u662f\u4ec0\u4e48\u3001\u5927\u6a21\u578b\u8c03\u7528\u8fc7\u7a0b\u4e2d Token \u5982\u4f55\u6d88\u8017\u3001\u4ee5\u53ca\u4ee5\u5343\u95ee\uff08Qwen\uff09\u7cfb\u5217\u4e3a\u4f8b\u7684\u8ba1\u8d39\u5168\u8c8c<\/p>\n<\/div>\n\n\n<div class=\"card\">\n <h2>\u4e00\u3001Token \u662f\u4ec0\u4e48\uff1f<\/h2>\n <p><strong>Token<\/strong> \u662f\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u5904\u7406\u6587\u672c\u7684<strong>\u6700\u5c0f\u57fa\u672c\u5355\u4f4d<\/strong>\u3002\u53ef\u4ee5\u628a\u5b83\u7406\u89e3\u4e3a\u6a21\u578b”\u9605\u8bfb”\u548c”\u5199\u4f5c”\u65f6\u4f7f\u7528\u7684”\u539f\u5b50\u8bcd\u6c47”\u2014\u2014\u6a21\u578b\u4e0d\u4f1a\u76f4\u63a5\u5904\u7406\u6c49\u5b57\u6216\u82f1\u6587\u5b57\u6bcd\uff0c\u800c\u662f\u5c06\u5b83\u4eec\u62c6\u89e3\u4e3a Token \u5e8f\u5217\u3002<\/p>\n\n <h3>Token \u5207\u5206\u793a\u4f8b<\/h3>\n <div class=\"table-wrap\">\n <table>\n <tr><th>\u539f\u59cb\u6587\u672c<\/th><th>Token \u6570\u91cf<\/th><th>Token \u5207\u5206\u7ed3\u679c<\/th><\/tr>\n <tr><td><code>\u4eba\u5de5\u667a\u80fd<\/code>\uff08\u4e2d\u6587\uff09<\/td><td>\u7ea6 2-4 \u4e2a<\/td><td><code>\u4eba\u5de5<\/code> + <code>\u667a\u80fd<\/code><\/td><\/tr>\n <tr><td><code>Hello World<\/code>\uff08\u82f1\u6587\uff09<\/td><td>\u7ea6 2-3 \u4e2a<\/td><td><code>Hello<\/code> + <code> World<\/code><\/td><\/tr>\n <tr><td><code>transformer<\/code>\uff08\u82f1\u6587\u957f\u8bcd\uff09<\/td><td>\u7ea6 2-3 \u4e2a<\/td><td><code>trans<\/code> + <code>former<\/code><\/td><\/tr>\n <tr><td><code>123456<\/code>\uff08\u6570\u5b57\uff09<\/td><td>1-3 \u4e2a<\/td><td><code>123<\/code> + <code>456<\/code><\/td><\/tr>\n <tr><td><code>print(\"hello\")<\/code>\uff08\u4ee3\u7801\uff09<\/td><td>\u7ea6 5-6 \u4e2a<\/td><td><code>print<\/code> + <code>(<\/code> + <code>\"<\/code> + <code>hello<\/code> + <code>\"<\/code> + <code>)<\/code><\/td><\/tr>\n <\/table>\n <\/div>\n\n <div class=\"callout info\">\n <strong>\u7ecf\u9a8c\u6cd5\u5219<\/strong>\n 1 \u4e2a\u4e2d\u6587\u5b57\u7b26 ≈ 1.5–2 \u4e2a Token | 1 \u4e2a\u82f1\u6587\u5355\u8bcd ≈ 1–1.5 \u4e2a Token | 1000 \u4e2a Token ≈ \u7ea6 750 \u4e2a\u82f1\u6587\u5355\u8bcd ≈ \u7ea6 500 \u4e2a\u4e2d\u6587\u6c49\u5b57\n <\/div>\n\n <h3>\u4e3a\u4ec0\u4e48\u9700\u8981 Token\uff1f<\/h3>\n <ul>\n <li><strong>\u6a21\u578b\u53ea\u80fd\u5904\u7406\u6570\u5b57<\/strong>\uff1aLLM \u5185\u90e8\u662f\u795e\u7ecf\u7f51\u7edc\uff0c\u6240\u6709\u8f93\u5165\u5fc5\u987b\u5148\u8f6c\u5316\u4e3a\u6570\u503c\u5411\u91cf\uff0cToken \u5c31\u662f\u6587\u672c\u5230\u6570\u5b57\u7684\u6865\u6881<\/li>\n <li><strong>\u8ba1\u8d39\u7684\u57fa\u7840\u5355\u5143<\/strong>\uff1aAPI \u6309 Token \u6570\u91cf\u6536\u8d39\uff0c\u7528\u591a\u5c11\u4ed8\u591a\u5c11\uff0c\u516c\u5e73\u900f\u660e<\/li>\n <li><strong>\u4e0a\u4e0b\u6587\u7a97\u53e3\u7684\u5ea6\u91cf<\/strong>\uff1a\u6a21\u578b\u80fd\u4e00\u6b21”\u8bb0\u4f4f”\u7684\u5185\u5bb9\u4e0a\u9650\u4ee5 Token \u6570\u8861\u91cf\uff08\u5982 128K Token\uff09<\/li>\n <\/ul>\n<\/div>\n\n\n<div class=\"card\">\n <h2>\u4e8c\u3001LLM \u8c03\u7528\u8fc7\u7a0b\u4e2d Token \u5982\u4f55\u6d88\u8017\uff1f<\/h2>\n <p>\u4e00\u6b21\u5b8c\u6574\u7684 LLM \u8c03\u7528\uff0cToken \u6d88\u8017\u53d1\u751f\u5728\u4e09\u4e2a\u73af\u8282\uff1a<\/p>\n\n <h3>2.1 \u7cfb\u7edf\u63d0\u793a\u8bcd\uff08System Prompt\uff09<\/h3>\n <p>\u5f00\u53d1\u8005\u9884\u8bbe\u7684”\u89d2\u8272\u8bbe\u5b9a”\uff0c\u5982”\u4f60\u662f\u4e00\u4e2a\u4e13\u4e1a\u7684\u6cd5\u5f8b\u987e\u95ee”\u3002\u6bcf\u6b21\u8c03\u7528\u90fd\u4f1a<strong>\u91cd\u590d\u8ba1\u5165<\/strong>\u8f93\u5165 Token\u3002\u7cfb\u7edf\u63d0\u793a\u8bcd\u8d8a\u957f\uff0c\u6bcf\u6b21\u5bf9\u8bdd\u7684\u56fa\u5b9a\u6210\u672c\u8d8a\u9ad8\u3002<\/p>\n <pre><code>system_prompt = \"\u4f60\u662f\u4e00\u4e2a\u5730\u8d28\u5efa\u6a21\u4e13\u5bb6\uff0c\u8bf7\u7528\u4e13\u4e1a\u672f\u8bed\u56de\u7b54\u3002\"\n# \u8fd9\u6761\u63d0\u793a\u8bcd\u7ea6 10-15 \u4e2a Token\uff0c\u6bcf\u6b21\u5bf9\u8bdd\u90fd\u4f1a\u6d88\u8017<\/code><\/pre>\n\n <h3>2.2 \u8f93\u5165 Token\uff08Input \/ Prompt Tokens\uff09<\/h3>\n <p>\u7528\u6237\u63d0\u51fa\u7684\u95ee\u9898\u6216\u7ed9\u6a21\u578b\u7684\u6240\u6709\u6587\u672c\u5185\u5bb9\uff0c\u5305\u62ec\uff1a<\/p>\n <ul>\n <li>\u7528\u6237\u6d88\u606f\u6b63\u6587<\/li>\n <li>\u5bf9\u8bdd\u5386\u53f2\uff08\u591a\u8f6e\u5bf9\u8bdd\u65f6\uff0c\u4e4b\u524d\u7684\u95ee\u7b54\u4f1a\u91cd\u65b0\u53d1\u9001\uff09<\/li>\n <li>\u4e0a\u4f20\u7684\u6587\u4ef6\u5185\u5bb9\uff08\u5982 PDF \u6458\u8981\u3001\u4ee3\u7801\u6587\u4ef6\uff09<\/li>\n <li>\u5de5\u5177\u8c03\u7528\u7ed3\u679c\uff08Function Calling \u7684\u8fd4\u56de\u503c\uff09<\/li>\n <\/ul>\n\n <div class=\"callout warn\">\n <strong>\u591a\u8f6e\u5bf9\u8bdd\u7684\u9690\u85cf\u6210\u672c<\/strong>\n \u6bcf\u6b21\u65b0\u5bf9\u8bdd\uff0c<strong>\u5168\u90e8\u5386\u53f2\u6d88\u606f<\/strong>\u4f1a\u88ab\u91cd\u65b0\u53d1\u9001\u3002\u7b2c 10 \u8f6e\u5bf9\u8bdd\u65f6\uff0c\u524d 9 \u8f6e\u7684\u95ee\u7b54\u5df2\u7d2f\u8ba1\u6210\u4e3a\u5927\u91cf\u8f93\u5165 Token\u3002\u8fd9\u5c31\u662f\u4e3a\u4ec0\u4e48\u957f\u5bf9\u8bdd\u8d8a\u6765\u8d8a\u8d35\u3002\n <\/div>\n\n <h3>2.3 \u8f93\u51fa Token\uff08Output \/ Completion Tokens\uff09<\/h3>\n <p>\u6a21\u578b\u751f\u6210\u7684\u56de\u590d\u6587\u672c\u3002\u8f93\u51fa Token \u6309\u5b9e\u9645\u751f\u6210\u6570\u91cf\u8ba1\u7b97\uff0c\u53d7 <code>max_tokens<\/code> \u53c2\u6570\u9650\u5236\u3002<\/p>\n\n <h3>2.4 \u5b8c\u6574\u6d88\u8017\u6d41\u7a0b<\/h3>\n <pre><code>\u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\n\u2502 \u4e00\u6b21 API \u8c03\u7528\u7684 Token \u8d26\u5355 \u2502\n\u251c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2524\n\u2502 \u8f93\u5165 Token\uff08\u4f60\u53d1\u7ed9\u6a21\u578b\u7684\uff09 \u2502\n\u2502 \u251c\u2500\u2500 \u7cfb\u7edf\u63d0\u793a\u8bcd\uff08System Prompt\uff09 \u2502\n\u2502 \u251c\u2500\u2500 \u7528\u6237\u6d88\u606f\uff08\u5f53\u524d\u95ee\u9898\uff09 \u2502\n\u2502 \u251c\u2500\u2500 \u5bf9\u8bdd\u5386\u53f2\uff08\u4e4b\u524d\u7684\u591a\u8f6e\u95ee\u7b54\uff09 \u2502\n\u2502 \u2514\u2500\u2500 \u5de5\u5177\/\u6587\u4ef6\u4e0a\u4e0b\u6587\uff08\u5982\u6709\uff09 \u2502\n\u251c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2524\n\u2502 \u8f93\u51fa Token\uff08\u6a21\u578b\u751f\u6210\u7684\u56de\u590d\uff09 \u2502\n\u2502 \u2514\u2500\u2500 \u6a21\u578b\u56de\u7b54\u7684\u5168\u90e8\u5185\u5bb9 \u2502\n\u251c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2524\n\u2502 \u603b\u6d88\u8017 = \u8f93\u5165 Token + \u8f93\u51fa Token \u2502\n\u2502 \u603b\u8d39\u7528 = \u8f93\u5165 Token \u00d7 \u8f93\u5165\u5355\u4ef7 + \u8f93\u51fa Token \u00d7 \u8f93\u51fa\u5355\u4ef7 \u2502\n\u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518<\/code><\/pre>\n<\/div>\n\n\n<div class=\"card\">\n <h2>\u4e09\u3001Token \u8ba1\u8d39\u7684\u57fa\u672c\u516c\u5f0f<\/h2>\n <div class=\"cmd\"><span class=\"p\">$<\/span><span class=\"c\">\u603b\u8d39\u7528 = (\u8f93\u5165Token\u6570 × \u8f93\u5165\u5355\u4ef7) + (\u8f93\u51faToken\u6570 × \u8f93\u51fa\u5355\u4ef7)<\/span><\/div>\n\n <h3>\u8ba1\u7b97\u793a\u4f8b<\/h3>\n <p>\u5047\u8bbe\u4f7f\u7528\u5343\u95ee Qwen-Long \u6a21\u578b\uff08\u8f93\u5165 0.5 \u5143\/\u767e\u4e07Token\uff0c\u8f93\u51fa 2 \u5143\/\u767e\u4e07Token\uff09\uff1a<\/p>\n <pre><code># \u7528\u6237\u63d0\u95ee\uff1a\"\u8bf7\u89e3\u91ca\u4ec0\u4e48\u662f Token\"\uff08\u7ea6 300 Token\uff09\n# \u6a21\u578b\u56de\u590d\u4e86\u8be6\u7ec6\u89e3\u91ca\uff08\u7ea6 900 Token\uff09\n\n\u8f93\u5165\u8d39\u7528 = 300 \/ 1,000,000 \u00d7 0.5 = \u00a50.00015\n\u8f93\u51fa\u8d39\u7528 = 900 \/ 1,000,000 \u00d7 2.0 = \u00a50.0018\n\u603b\u8d39\u7528 = \u00a50.00195\uff08\u7ea6 0.2 \u5206\u94b1\uff09\n\n# \u4e00\u5929\u8c03\u7528 1000 \u6b21 \u2248 \u00a51.95\n# \u4e00\u5929\u8c03\u7528 10 \u4e07\u6b21 \u2248 \u00a5195<\/code><\/pre>\n\n <div class=\"callout tip\">\n <strong>\u5173\u952e\u8ba4\u77e5<\/strong>\n \u8f93\u51fa Token \u7684\u5355\u4ef7\u901a\u5e38\u662f\u8f93\u5165 Token \u7684 <strong>3–10 \u500d<\/strong>\u3002\u56e0\u6b64\uff0c\u63a7\u5236\u6a21\u578b\u56de\u590d\u957f\u5ea6\uff08\u8bbe\u7f6e\u5408\u7406\u7684 <code>max_tokens<\/code>\uff09\u6bd4\u7cbe\u7b80\u8f93\u5165\u66f4\u7701\u94b1\u3002\n <\/div>\n\n <h3>\u591a\u8f6e\u5bf9\u8bdd\u7684 Token \u7d2f\u79ef<\/h3>\n <pre><code>\u7b2c 1 \u8f6e\uff1a\u8f93\u5165 500 Token \u2192 \u8f93\u51fa 800 Token \u2192 \u7d2f\u8ba1 1300\n\u7b2c 2 \u8f6e\uff1a\u8f93\u5165 500+1300=1800 Token \u2192 \u8f93\u51fa 600 Token \u2192 \u7d2f\u8ba1 3700\n\u7b2c 3 \u8f6e\uff1a\u8f93\u5165 500+3700=4200 Token \u2192 \u8f93\u51fa 500 Token \u2192 \u7d2f\u8ba1 8400\n...\n\u7b2c 10 \u8f6e\uff1a\u8f93\u5165\u53ef\u80fd\u5df2\u8d85\u8fc7 20000 Token\n\n# \u6bcf\u6b21\u5bf9\u8bdd\u589e\u957f\uff0c\u8f93\u5165 Token \u5448\u7ebf\u6027\u589e\u957f<\/code><\/pre>\n<\/div>\n\n\n<div class=\"card\">\n <h2>\u56db\u3001\u5343\u95ee\uff08Qwen\uff09\u7cfb\u5217\u8ba1\u8d39\u8be6\u89e3<\/h2>\n <p>\u963f\u91cc\u4e91\u5343\u95ee\u7cfb\u5217\u6309 Token \u8ba1\u8d39\uff0c\u4ef7\u683c\u56e0\u6a21\u578b\u3001\u8f93\u5165\u957f\u5ea6\u3001\u90e8\u7f72\u533a\u57df\u548c\u8c03\u7528\u6a21\u5f0f\u800c\u5f02\u3002\u4ee5\u4e0b\u4e3a 2026 \u5e74<strong>\u4e2d\u56fd\u5185\u5730\uff08\u5317\u4eac\u5730\u57df\uff09<\/strong>\u4e3b\u8981\u6a21\u578b\u4ef7\u683c\u53c2\u8003\uff1a<\/p>\n\n <div class=\"table-wrap\">\n <table>\n <thead>\n <tr><th>\u6a21\u578b<\/th><th>\u8f93\u5165\u5355\u4ef7\uff08\/\u767e\u4e07Token\uff09<\/th><th>\u8f93\u51fa\u5355\u4ef7\uff08\/\u767e\u4e07Token\uff09<\/th><th>\u7279\u70b9<\/th><\/tr>\n <\/thead>\n <tbody>\n <tr><td><strong>Qwen-Long<\/strong><\/td><td><strong>\u00a50.5<\/strong><\/td><td><strong>\u00a52.0<\/strong><\/td><td>\u957f\u6587\u672c\u4e13\u7528\uff0c\u6027\u4ef7\u6bd4\u6700\u9ad8<\/td><\/tr>\n <tr><td><strong>Qwen3.5-Plus<\/strong><\/td><td><strong>\u00a50.8<\/strong><\/td><td><strong>\u00a54.8<\/strong><\/td><td>\u4e3b\u529b\u6a21\u578b\uff0c\u591a\u6a21\u6001\uff0c0-128K \u4e0a\u4e0b\u6587<\/td><\/tr>\n <tr><td>Qwen3.5-Plus (128K-256K)<\/td><td>\u00a52.0<\/td><td>\u00a512.0<\/td><td>\u8d85\u957f\u8f93\u5165\u533a\u95f4<\/td><\/tr>\n <tr><td>Qwen3.5-Plus (256K-1M)<\/td><td>\u00a54.0<\/td><td>\u00a524.0<\/td><td>\u767e\u4e07\u7ea7\u4e0a\u4e0b\u6587\u533a\u95f4<\/td><\/tr>\n <tr><td><strong>Qwen3-Max<\/strong><\/td><td><strong>\u00a52.5<\/strong><\/td><td><strong>\u00a510.0<\/strong><\/td><td>\u65d7\u8230\u6a21\u578b\uff0c0-32K \u4e0a\u4e0b\u6587<\/td><\/tr>\n <tr><td>Qwen3-Max (32K-128K)<\/td><td>\u00a54.0<\/td><td>\u00a516.0<\/td><td>\u4e2d\u957f\u6587\u672c\u533a\u95f4<\/td><\/tr>\n <tr><td>Qwen3-Max (128K-252K)<\/td><td>\u00a57.0<\/td><td>\u00a528.0<\/td><td>\u8d85\u957f\u6587\u672c\u533a\u95f4<\/td><\/tr>\n <tr><td><strong>Qwen3-Omni-Flash<\/strong><\/td><td>\u00a51.8 <small>(\u6587\u672c)<\/small><\/td><td>\u00a56.9<\/td><td>\u591a\u6a21\u6001\uff08\u542b\u97f3\u9891\/\u56fe\u7247\/\u89c6\u9891\u8f93\u5165\uff09<\/td><\/tr>\n <tr><td>Qwen3-Omni-Flash <small>(\u97f3\u9891\u8f93\u5165)<\/small><\/td><td>\u00a515.8<\/td><td>\u00a512.7<\/td><td>\u97f3\u9891\u8f93\u5165\u4ef7\u683c\u66f4\u9ad8<\/td><\/tr>\n <\/tbody>\n <\/table>\n <\/div>\n\n <h3>\u7701\u94b1\u6280\u5de7<\/h3>\n <div class=\"compare-grid\">\n <div class=\"compare-box\">\n <div class=\"box-title\">Batch \u8c03\u7528\uff08\u534a\u4ef7\uff09<\/div>\n <p style=\"font-size:.85rem;\">\u4e0d\u8981\u6c42\u5b9e\u65f6\u54cd\u5e94\u7684\u4efb\u52a1\u53ef\u63d0\u4ea4\u6279\u91cf\u8bf7\u6c42\uff0c\u8f93\u5165\u548c\u8f93\u51fa\u5355\u4ef7\u5747\u4e3a\u5b9e\u65f6\u63a8\u7406\u7684 <strong>50%<\/strong>\u3002\u9002\u5408\u79bb\u7ebf\u5206\u6790\u3001\u6570\u636e\u6807\u6ce8\u7b49\u573a\u666f\u3002<\/p>\n <\/div>\n <div class=\"compare-box\">\n <div class=\"box-title\">\u4e0a\u4e0b\u6587\u7f13\u5b58<\/div>\n <p style=\"font-size:.85rem;\">\u9ad8\u9891\u91cd\u590d\u7684\u7cfb\u7edf\u63d0\u793a\u8bcd\u6216\u6587\u6863\u53ef\u7f13\u5b58\uff0c\u547d\u4e2d\u7f13\u5b58\u65f6<strong>\u4ec5\u8f93\u5165 Token \u4eab\u6298\u6263<\/strong>\u3002\u9002\u5408\u56fa\u5b9a Prompt \u7684\u591a\u6b21\u8c03\u7528\u3002<\/p>\n <\/div>\n <div class=\"compare-box\">\n <div class=\"box-title\">\u514d\u8d39\u989d\u5ea6<\/div>\n <p style=\"font-size:.85rem;\">\u65b0\u7528\u6237\u5f00\u901a\u963f\u91cc\u4e91\u767e\u70bc\u5e73\u53f0\u5373\u9001\u5404\u6a21\u578b <strong>100 \u4e07 Token<\/strong> \u514d\u8d39\u989d\u5ea6\uff0890 \u5929\u6709\u6548\uff09\uff0c\u8fd8\u53ef\u9886\u53d6\u8d85 7000 \u4e07 Token\u3002<\/p>\n <\/div>\n <div class=\"compare-box\">\n <div class=\"box-title\">\u5408\u7406\u9009\u62e9\u6a21\u578b<\/div>\n <p style=\"font-size:.85rem;\">\u7b80\u5355\u4efb\u52a1\u7528 Qwen-Long\uff08\u6700\u4f4e\u4ef7\uff09\uff0c\u590d\u6742\u63a8\u7406\u7528 Qwen-Max\u3002\u4e0d\u8981\u7528\u65d7\u8230\u6a21\u578b\u5904\u7406\u65e5\u5e38\u95f2\u804a\u3002<\/p>\n <\/div>\n <\/div>\n\n <div class=\"callout info\">\n <strong>\u4ef7\u683c\u53c2\u8003<\/strong>\n \u4ee5\u4e0a\u4ef7\u683c\u4e3a\u963f\u91cc\u4e91\u767e\u70bc\u5e73\u53f0\u4e2d\u56fd\u5185\u5730\uff08\u5317\u4eac\u5730\u57df\uff09\u516c\u5f00\u5b9a\u4ef7\uff0c\u5b9e\u9645\u4ef7\u683c\u4ee5\u5b98\u65b9\u6700\u65b0\u516c\u544a\u4e3a\u51c6\u3002\u56fd\u9645\u90e8\u7f72\uff08\u65b0\u52a0\u5761\uff09\u4ef7\u683c\u7ea6\u4e3a\u5185\u5730\u7684 3–5 \u500d\u3002<br>\n \u5b98\u65b9\u5b9a\u4ef7\u9875\uff1a<a href=\"https:\/\/www.aliyun.com\/product\/tongyi\" target=\"_blank\">aliyun.com\/product\/tongyi<\/a> | \u767e\u70bc\u5e73\u53f0\uff1a<a href=\"https:\/\/www.aliyun.com\/product\/bailian\" target=\"_blank\">aliyun.com\/product\/bailian<\/a>\n <\/div>\n<\/div>\n\n\n<div class=\"card\">\n <h2>\u4e94\u3001\u5f71\u54cd Token \u6d88\u8017\u7684\u989d\u5916\u56e0\u7d20<\/h2>\n <div class=\"table-wrap\">\n <table>\n <tr><th>\u56e0\u7d20<\/th><th>\u5f71\u54cd<\/th><th>\u8bf4\u660e<\/th><\/tr>\n <tr><td><strong>\u601d\u8003\u6a21\u5f0f\uff08Thinking\uff09<\/strong><\/td><td>\u8f93\u51fa Token \u5927\u5e45\u589e\u52a0<\/td><td>\u6a21\u578b\u5185\u90e8\u63a8\u7406\u8fc7\u7a0b\u4e5f\u8ba1\u5165\u8f93\u51fa Token\uff0c\u53ef\u80fd\u7ffb\u500d\u751a\u81f3\u66f4\u591a<\/td><\/tr>\n <tr><td><strong>Function Calling<\/strong><\/td><td>\u8f93\u5165 Token \u589e\u52a0<\/td><td>\u5de5\u5177\u5b9a\u4e49\u548c\u8fd4\u56de\u7ed3\u679c\u90fd\u8ba1\u5165\u8f93\u5165 Token<\/td><\/tr>\n <tr><td><strong>\u591a\u6a21\u6001\u8f93\u5165<\/strong><\/td><td>\u5355\u4ef7\u66f4\u9ad8<\/td><td>\u56fe\u7247\u3001\u97f3\u9891\u3001\u89c6\u9891\u6309\u5404\u81ea\u5355\u4ef7\u8ba1\u8d39\uff0c\u4e14\u901a\u5e38\u8fdc\u9ad8\u4e8e\u7eaf\u6587\u672c<\/td><\/tr>\n <tr><td><strong>max_tokens \u8bbe\u7f6e<\/strong><\/td><td>\u9650\u5236\u8f93\u51fa\u4e0a\u9650<\/td><td>\u8bbe\u7f6e\u8fc7\u5927\u4f1a\u5bfc\u81f4\u975e\u5fc5\u8981\u7684\u9ad8\u6d88\u8017\uff0c\u5efa\u8bae\u6839\u636e\u5b9e\u9645\u9700\u6c42\u8bbe\u5b9a\u5408\u7406\u503c<\/td><\/tr>\n <tr><td><strong>\u6d41\u5f0f\u8f93\u51fa\uff08Streaming\uff09<\/strong><\/td><td>\u4e0d\u51cf\u5c11\u603b\u6d88\u8017<\/td><td>\u4f46\u53ef\u4e2d\u9014\u505c\u6b62\u751f\u6210\uff0c\u907f\u514d\u4e0d\u5fc5\u8981\u8f93\u51fa<\/td><\/tr>\n <\/table>\n <\/div>\n\n <h3>\u601d\u8003\u6a21\u5f0f\u7684 Token \u5f00\u9500<\/h3>\n <pre><code># \u666e\u901a\u6a21\u5f0f\n\u8f93\u5165 500 Token \u2192 \u8f93\u51fa 800 Token \u2192 \u603b\u6d88\u8017 1300 Token\n\n# \u601d\u8003\u6a21\u5f0f\uff08\u5f00\u542f\u6df1\u5ea6\u63a8\u7406\uff09\n\u8f93\u5165 500 Token \u2192 \u601d\u8003\u8fc7\u7a0b 1200 Token + \u6700\u7ec8\u56de\u7b54 600 Token\n\u2192 \u8f93\u51fa 1800 Token \u2192 \u603b\u6d88\u8017 2300 Token\n# \u8d39\u7528\u53ef\u80fd\u662f\u666e\u901a\u6a21\u5f0f\u7684 2-3 \u500d<\/code><\/pre>\n<\/div>\n\n\n<div class=\"card\">\n <h2>\u516d\u3001\u5982\u4f55\u8ba1\u7b97\u548c\u76d1\u63a7 Token \u6d88\u8017<\/h2>\n <h3>6.1 API \u8fd4\u56de\u503c\u4e2d\u67e5\u770b<\/h3>\n <p>\u6bcf\u6b21\u8c03\u7528 API \u540e\uff0c\u54cd\u5e94\u4e2d\u4f1a\u5305\u542b <code>usage<\/code> \u5b57\u6bb5\uff0c\u76f4\u63a5\u663e\u793a\u672c\u6b21\u6d88\u8017\uff1a<\/p>\n <pre><code>{\n \"usage\": {\n \"prompt_tokens\": 350, \/\/ \u8f93\u5165 Token\n \"completion_tokens\": 820, \/\/ \u8f93\u51fa Token\n \"total_tokens\": 1170 \/\/ \u603b\u6d88\u8017\n }\n}<\/code><\/pre>\n\n <h3>6.2 \u4f7f\u7528 Tokenizer \u9884\u5148\u4f30\u7b97<\/h3>\n <pre><code>from transformers import AutoTokenizer\n\n# \u52a0\u8f7d\u5343\u95ee\u7684 Tokenizer\ntokenizer = AutoTokenizer.from_pretrained(\n \"Qwen\/Qwen2.5-7B-Instruct\", trust_remote_code=True\n)\n\n# \u4f30\u7b97\u8f93\u5165 Token \u6570\ntext = \"\u8bf7\u7528\u901a\u4fd7\u7684\u8bed\u8a00\u89e3\u91ca\u4ec0\u4e48\u662f Token\"\ntokens = tokenizer.encode(text)\nprint(f\"\u8f93\u5165 Token \u6570: {len(tokens)}\")\n# \u8f93\u51fa: \u8f93\u5165 Token \u6570: ~15-20<\/code><\/pre>\n\n <h3>6.3 \u6210\u672c\u8ba1\u7b97\u51fd\u6570<\/h3>\n <pre><code>def calc_cost(input_tokens, output_tokens,\n input_price_per_million=0.8,\n output_price_per_million=4.8):\n \"\"\"\u8ba1\u7b97\u5343\u95ee 3.5-Plus \u7684\u5355\u6b21\u8c03\u7528\u8d39\u7528\"\"\"\n input_cost = (input_tokens \/ 1_000_000) * input_price_per_million\n output_cost = (output_tokens \/ 1_000_000) * output_price_per_million\n return round(input_cost + output_cost, 6)\n\n# \u793a\u4f8b\ncost = calc_cost(500, 1200)\nprint(f\"\u8d39\u7528: \u00a5{cost}\") # \u8d39\u7528: \u00a50.00616<\/code><\/pre>\n<\/div>\n\n\n<div class=\"card\">\n <h2>\u4e03\u3001\u7701\u94b1\u6700\u4f73\u5b9e\u8df5<\/h2>\n <ol>\n <li><strong>\u7cbe\u7b80\u7cfb\u7edf\u63d0\u793a\u8bcd<\/strong>\uff1a\u53bb\u6389\u4e0d\u5fc5\u8981\u7684\u80cc\u666f\u8bf4\u660e\uff0c\u6bcf\u6761\u591a\u51fa\u7684 Token \u90fd\u4f1a\u5728\u6bcf\u6b21\u5bf9\u8bdd\u4e2d\u91cd\u590d\u6d88\u8017<\/li>\n <li><strong>\u63a7\u5236\u5bf9\u8bdd\u8f6e\u6b21<\/strong>\uff1a\u957f\u5bf9\u8bdd\u7684\u5386\u53f2\u7d2f\u79ef\u662f\u9690\u85cf\u7684\u6210\u672c\u5927\u5934\uff0c\u5fc5\u8981\u65f6\u53ef\u5b9a\u671f”\u91cd\u7f6e\u5bf9\u8bdd”<\/li>\n <li><strong>\u8bbe\u7f6e\u5408\u7406\u7684 max_tokens<\/strong>\uff1a\u4e0d\u8981\u9ed8\u8ba4\u8bbe 4096 \u6216 8192\uff0c\u6309\u5b9e\u9645\u56de\u590d\u957f\u5ea6\u9700\u6c42\u8bbe\u5b9a<\/li>\n <li><strong>\u7b80\u5355\u4efb\u52a1\u964d\u7ea7\u6a21\u578b<\/strong>\uff1a\u65e5\u5e38\u95ee\u7b54\u7528 Qwen-Long\uff080.5 \u5143\/\u767e\u4e07Token\uff09\uff0c\u590d\u6742\u63a8\u7406\u624d\u7528 Qwen-Max<\/li>\n <li><strong>\u6279\u91cf\u4efb\u52a1\u7528 Batch \u6a21\u5f0f<\/strong>\uff1a\u79bb\u7ebf\u5904\u7406\u5927\u6279\u91cf\u6570\u636e\u65f6\uff0cBatch \u6a21\u5f0f\u53ef\u8282\u7701 50% \u8d39\u7528<\/li>\n <li><strong>\u542f\u7528\u6d41\u5f0f\u8f93\u51fa<\/strong>\uff1a\u7528\u6237\u53ef\u4e2d\u9014\u505c\u6b62\u751f\u6210\uff0c\u907f\u514d\u6a21\u578b\u4ea7\u51fa\u4e0d\u9700\u8981\u7684\u5197\u957f\u56de\u590d<\/li>\n <li><strong>\u9ad8\u9891\u95ee\u7b54\u505a\u7f13\u5b58<\/strong>\uff1a\u76f8\u540c\u95ee\u9898\u76f4\u63a5\u8fd4\u56de\u7f13\u5b58\u7ed3\u679c\uff0c\u96f6 Token \u6d88\u8017<\/li>\n <li><strong>\u5229\u7528\u514d\u8d39\u989d\u5ea6<\/strong>\uff1a\u5f00\u901a\u767e\u70bc\u5e73\u53f0\u9886\u53d6 7000 \u4e07+ Token \u514d\u8d39\u989d\u5ea6\uff0c\u5148\u8bd5\u7528\u518d\u4ed8\u8d39<\/li>\n <\/ol>\n<\/div>\n\n<div class=\"summary-box\">\n <h2>\u6838\u5fc3\u8981\u70b9\u603b\u7ed3<\/h2>\n <div class=\"s-row\"><div class=\"s-label\">Token \u672c\u8d28<\/div><div class=\"s-desc\">LLM \u5904\u7406\u6587\u672c\u7684\u6700\u5c0f\u5355\u4f4d\uff0c\u4e2d\u6587\u7ea6 1-2 \u5b57\/Token\uff0c\u82f1\u6587\u7ea6 1 \u8bcd\/Token<\/div><\/div>\n <div class=\"s-row\"><div class=\"s-label\">\u6d88\u8017\u516c\u5f0f<\/div><div class=\"s-desc\">\u603b\u6d88\u8017 = \u8f93\u5165 Token + \u8f93\u51fa Token\uff1b\u603b\u8d39\u7528 = \u8f93\u5165 \u00d7 \u8f93\u5165\u5355\u4ef7 + \u8f93\u51fa \u00d7 \u8f93\u51fa\u5355\u4ef7<\/div><\/div>\n <div class=\"s-row\"><div class=\"s-label\">\u8f93\u5165\u6765\u6e90<\/div><div class=\"s-desc\">\u7cfb\u7edf\u63d0\u793a\u8bcd + \u7528\u6237\u6d88\u606f + \u5bf9\u8bdd\u5386\u53f2 + \u5de5\u5177\/\u6587\u4ef6\u4e0a\u4e0b\u6587<\/div><\/div>\n <div class=\"s-row\"><div class=\"s-label\">\u8f93\u51fa\u7279\u70b9<\/div><div class=\"s-desc\">\u5355\u4ef7\u901a\u5e38\u662f\u8f93\u5165\u7684 3-10 \u500d\uff0c\u601d\u8003\u6a21\u5f0f\u989d\u5916\u589e\u52a0\u63a8\u7406 Token<\/div><\/div>\n <div class=\"s-row\"><div class=\"s-label\">\u5343\u95ee\u4ef7\u683c<\/div><div class=\"s-desc\">Long \u6700\u4f4e \u00a50.5\/\u767e\u4e07Token \u8f93\u5165\uff1bPlus \u4e3b\u529b \u00a50.8 \u8f93\u5165\uff1bMax \u65d7\u8230 \u00a52.5 \u8f93\u5165<\/div><\/div>\n <div class=\"s-row\"><div class=\"s-label\">\u7701\u94b1\u6838\u5fc3<\/div><div class=\"s-desc\">\u7cbe\u7b80 Prompt + \u63a7\u5236\u5bf9\u8bdd\u8f6e\u6b21 + \u9009\u5bf9\u6a21\u578b + Batch \u534a\u4ef7 + \u5229\u7528\u514d\u8d39\u989d\u5ea6<\/div><\/div>\n<\/div>\n\n<\/div>\n\n\n<div class=\"lang-section\" id=\"lang-en\">\n\n<div class=\"page-header\">\n <span class=\"badge\">Token Guide<\/span>\n <h1>Complete Token Guide: LLM Consumption & Pricing<\/h1>\n <p>Understand what tokens are, how they are consumed during LLM calls, and a complete pricing overview using the Qwen model series as an example<\/p>\n<\/div>\n\n\n<div class=\"card\">\n <h2>1. What Is a Token?<\/h2>\n <p>A <strong>Token<\/strong> is the <strong>smallest basic unit<\/strong> that a Large Language Model (LLM) uses to process text. Think of it as the “atomic vocabulary” the model uses to “read” and “write” \u2014 the model doesn’t process Chinese characters or English letters directly, but breaks them into token sequences.<\/p>\n\n <h3>Tokenization Examples<\/h3>\n <div class=\"table-wrap\">\n <table>\n <tr><th>Original Text<\/th><th>Token Count<\/th><th>Token Breakdown<\/th><\/tr>\n <tr><td><code>artificial intelligence<\/code><\/td><td>~2-3<\/td><td><code>art<\/code> + <code>ificial<\/code> + <code> intelligence<\/code><\/td><\/tr>\n <tr><td><code>Hello World<\/code><\/td><td>~2-3<\/td><td><code>Hello<\/code> + <code> World<\/code><\/td><\/tr>\n <tr><td><code>transformer<\/code><\/td><td>~2-3<\/td><td><code>trans<\/code> + <code>former<\/code><\/td><\/tr>\n <tr><td><code>123456<\/code><\/td><td>1-3<\/td><td><code>123<\/code> + <code>456<\/code><\/td><\/tr>\n <tr><td><code>print(\"hello\")<\/code><\/td><td>~5-6<\/td><td><code>print<\/code> + <code>(<\/code> + <code>\"<\/code> + <code>hello<\/code> + <code>\"<\/code> + <code>)<\/code><\/td><\/tr>\n <\/table>\n <\/div>\n\n <div class=\"callout info\">\n <strong>Rule of Thumb<\/strong>\n 1 Chinese character ≈ 1.5–2 tokens | 1 English word ≈ 1–1.5 tokens | 1000 tokens ≈ ~750 English words ≈ ~500 Chinese characters\n <\/div>\n\n <h3>Why Tokens?<\/h3>\n <ul>\n <li><strong>Models only process numbers<\/strong>: LLMs are neural networks; all input must be converted to numerical vectors first. Tokens bridge text to numbers.<\/li>\n <li><strong>Unit of billing<\/strong>: APIs charge by token count \u2014 pay for what you use, transparent and fair.<\/li>\n <li><strong>Context window metric<\/strong>: The maximum amount of content a model can “remember” at once is measured in tokens (e.g., 128K tokens).<\/li>\n <\/ul>\n<\/div>\n\n\n<div class=\"card\">\n <h2>2. How Tokens Are Consumed During an LLM Call<\/h2>\n <p>A complete LLM call consumes tokens across three stages:<\/p>\n\n <h3>2.1 System Prompt<\/h3>\n <p>The developer-defined “role setting,” such as “You are a professional legal advisor.” It is <strong>re-counted every call<\/strong> into input tokens. The longer the system prompt, the higher the fixed cost per conversation.<\/p>\n <pre><code>system_prompt = \"You are a geological modeling expert. Use professional terms.\"\n# This prompt is ~10-15 tokens, consumed every conversation<\/code><\/pre>\n\n <h3>2.2 Input Tokens (Prompt Tokens)<\/h3>\n <p>All text content sent to the model, including:<\/p>\n <ul>\n <li>User message body<\/li>\n <li>Conversation history (prior Q&A rounds are re-sent in multi-turn chats)<\/li>\n <li>Uploaded file content (e.g., PDF excerpts, code files)<\/li>\n <li>Function calling results (tool return values)<\/li>\n <\/ul>\n\n <div class=\"callout warn\">\n <strong>Hidden cost of multi-turn conversations<\/strong>\n Every new turn, <strong>all historical messages<\/strong> are re-sent. By round 10, the previous 9 rounds have accumulated into a massive input token count. This is why long conversations get progressively more expensive.\n <\/div>\n\n <h3>2.3 Output Tokens (Completion Tokens)<\/h3>\n <p>The text the model generates in response. Output tokens are counted based on actual generation, limited by the <code>max_tokens<\/code> parameter.<\/p>\n\n <h3>2.4 Complete Consumption Flow<\/h3>\n <pre><code>\u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\n\u2502 Token Bill for One API Call \u2502\n\u251c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2524\n\u2502 Input Tokens (what you send to the model) \u2502\n\u2502 \u251c\u2500\u2500 System Prompt \u2502\n\u2502 \u251c\u2500\u2500 User Message (current question) \u2502\n\u2502 \u251c\u2500\u2500 Conversation History (prior Q&A rounds) \u2502\n\u2502 \u2514\u2500\u2500 Tool\/File Context (if any) \u2502\n\u251c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2524\n\u2502 Output Tokens (model's response) \u2502\n\u2502 \u2514\u2500\u2500 Full content of model's answer \u2502\n\u251c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2524\n\u2502 Total = Input Tokens + Output Tokens \u2502\n\u2502 Cost = Input \u00d7 Input Price + Output \u00d7 Output Price\u2502\n\u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518<\/code><\/pre>\n<\/div>\n\n\n<div class=\"card\">\n <h2>3. Basic Token Pricing Formula<\/h2>\n <div class=\"cmd\"><span class=\"p\">$<\/span><span class=\"c\">Total Cost = (Input Tokens × Input Price) + (Output Tokens × Output Price)<\/span><\/div>\n\n <h3>Calculation Example<\/h3>\n <p>Using Qwen-Long (input \u00a50.5\/million tokens, output \u00a52.0\/million tokens):<\/p>\n <pre><code># User asks: \"Explain what a token is\" (~300 tokens)\n# Model responds with detailed explanation (~900 tokens)\n\nInput cost = 300 \/ 1,000,000 \u00d7 0.5 = \u00a50.00015\nOutput cost = 900 \/ 1,000,000 \u00d7 2.0 = \u00a50.0018\nTotal cost = \u00a50.00195 (~0.2 Chinese cents)\n\n# 1,000 calls\/day \u2248 \u00a51.95\n# 100,000 calls\/day \u2248 \u00a5195<\/code><\/pre>\n\n <div class=\"callout tip\">\n <strong>Key Insight<\/strong>\n Output token pricing is typically <strong>3–10\u00d7<\/strong> higher than input. So controlling response length (setting a reasonable <code>max_tokens<\/code>) saves more than trimming input.\n <\/div>\n\n <h3>Multi-Turn Token Accumulation<\/h3>\n <pre><code>Round 1: Input 500 \u2192 Output 800 \u2192 Cumulative 1300\nRound 2: Input 500+1300=1800 \u2192 Output 600 \u2192 Cumulative 3700\nRound 3: Input 500+3700=4200 \u2192 Output 500 \u2192 Cumulative 8400\n...\nRound 10: Input may exceed 20,000 tokens\n\n# Input tokens grow linearly with each turn<\/code><\/pre>\n<\/div>\n\n\n<div class=\"card\">\n <h2>4. Qwen Series Pricing Breakdown<\/h2>\n <p>Alibaba Cloud’s Qwen series charges by token, with prices varying by model, input length, deployment region, and call mode. Below is the pricing reference for <strong>China mainland (Beijing region)<\/strong> as of 2026:<\/p>\n\n <div class=\"table-wrap\">\n <table>\n <thead>\n <tr><th>Model<\/th><th>Input (per M tokens)<\/th><th>Output (per M tokens)<\/th><th>Features<\/th><\/tr>\n <\/thead>\n <tbody>\n <tr><td><strong>Qwen-Long<\/strong><\/td><td><strong>\u00a50.5<\/strong><\/td><td><strong>\u00a52.0<\/strong><\/td><td>Long-text specialist, best value<\/td><\/tr>\n <tr><td><strong>Qwen3.5-Plus<\/strong><\/td><td><strong>\u00a50.8<\/strong><\/td><td><strong>\u00a54.8<\/strong><\/td><td>Main model, multimodal, 0-128K context<\/td><\/tr>\n <tr><td>Qwen3.5-Plus (128K-256K)<\/td><td>\u00a52.0<\/td><td>\u00a512.0<\/td><td>Extended input range<\/td><\/tr>\n <tr><td>Qwen3.5-Plus (256K-1M)<\/td><td>\u00a54.0<\/td><td>\u00a524.0<\/td><td>Million-token context range<\/td><\/tr>\n <tr><td><strong>Qwen3-Max<\/strong><\/td><td><strong>\u00a52.5<\/strong><\/td><td><strong>\u00a510.0<\/strong><\/td><td>Flagship model, 0-32K context<\/td><\/tr>\n <tr><td>Qwen3-Max (32K-128K)<\/td><td>\u00a54.0<\/td><td>\u00a516.0<\/td><td>Mid-long text range<\/td><\/tr>\n <tr><td>Qwen3-Max (128K-252K)<\/td><td>\u00a57.0<\/td><td>\u00a528.0<\/td><td>Ultra-long text range<\/td><\/tr>\n <tr><td><strong>Qwen3-Omni-Flash<\/strong><\/td><td>\u00a51.8 <small>(text)<\/small><\/td><td>\u00a56.9<\/td><td>Multimodal (audio\/image\/video input)<\/td><\/tr>\n <tr><td>Qwen3-Omni-Flash <small>(audio input)<\/small><\/td><td>\u00a515.8<\/td><td>\u00a512.7<\/td><td>Higher audio input pricing<\/td><\/tr>\n <\/tbody>\n <\/table>\n <\/div>\n\n <h3>Cost-Saving Tips<\/h3>\n <div class=\"compare-grid\">\n <div class=\"compare-box\">\n <div class=\"box-title\">Batch Calling (50% off)<\/div>\n <p style=\"font-size:.85rem;\">For non-real-time tasks, submit batch requests. Both input and output pricing is <strong>50%<\/strong> of real-time inference. Suitable for offline analysis and data labeling.<\/p>\n <\/div>\n <div class=\"compare-box\">\n <div class=\"box-title\">Context Caching<\/div>\n <p style=\"font-size:.85rem;\">High-frequency system prompts or documents can be cached. When the cache is hit, <strong>input tokens get a discount<\/strong>. Ideal for repeated calls with fixed prompts.<\/p>\n <\/div>\n <div class=\"compare-box\">\n <div class=\"box-title\">Free Quota<\/div>\n <p style=\"font-size:.85rem;\">New users of Alibaba Cloud Bailian Platform receive <strong>1 million tokens<\/strong> per model (90 days) plus over 70 million additional free tokens.<\/p>\n <\/div>\n <div class=\"compare-box\">\n <div class=\"box-title\">Choose the Right Model<\/div>\n <p style=\"font-size:.85rem;\">Use Qwen-Long (lowest price) for simple tasks; Qwen-Max for complex reasoning. Don’t use the flagship model for casual chat.<\/p>\n <\/div>\n <\/div>\n\n <div class=\"callout info\">\n <strong>Pricing Reference<\/strong>\n Prices above are public pricing for Alibaba Cloud Bailian Platform (Beijing region). Actual prices are subject to the latest official announcements. International deployment (Singapore) is approximately 3–5\u00d7 higher.<br>\n Official pricing: <a href=\"https:\/\/www.aliyun.com\/product\/tongyi\" target=\"_blank\">aliyun.com\/product\/tongyi<\/a> | Bailian Platform: <a href=\"https:\/\/www.aliyun.com\/product\/bailian\" target=\"_blank\">aliyun.com\/product\/bailian<\/a>\n <\/div>\n<\/div>\n\n\n<div class=\"card\">\n <h2>5. Additional Factors Affecting Token Consumption<\/h2>\n <div class=\"table-wrap\">\n <table>\n <tr><th>Factor<\/th><th>Impact<\/th><th>Description<\/th><\/tr>\n <tr><td><strong>Thinking Mode<\/strong><\/td><td>Output tokens increase significantly<\/td><td>Internal reasoning process is also counted as output, potentially doubling or more<\/td><\/tr>\n <tr><td><strong>Function Calling<\/strong><\/td><td>Input tokens increase<\/td><td>Tool definitions and return values are counted as input tokens<\/td><\/tr>\n <tr><td><strong>Multimodal Input<\/strong><\/td><td>Higher unit price<\/td><td>Images, audio, and video are billed at their own rates, usually much higher than plain text<\/td><\/tr>\n <tr><td><strong>max_tokens setting<\/strong><\/td><td>Caps output<\/td><td>Setting it too high leads to unnecessary consumption; set a reasonable value based on actual needs<\/td><\/tr>\n <tr><td><strong>Streaming<\/strong><\/td><td>Doesn’t reduce total<\/td><td>But allows mid-generation stopping to avoid unnecessary output<\/td><\/tr>\n <\/table>\n <\/div>\n\n <h3>Thinking Mode Token Overhead<\/h3>\n <pre><code># Normal mode\nInput 500 tokens \u2192 Output 800 tokens \u2192 Total 1300 tokens\n\n# Thinking mode (deep reasoning enabled)\nInput 500 tokens \u2192 Reasoning 1200 tokens + Final answer 600 tokens\n\u2192 Output 1800 tokens \u2192 Total 2300 tokens\n# Cost may be 2-3\u00d7 of normal mode<\/code><\/pre>\n<\/div>\n\n\n<div class=\"card\">\n <h2>6. How to Calculate and Monitor Token Consumption<\/h2>\n <h3>6.1 Check API Response<\/h3>\n <p>Every API response includes a <code>usage<\/code> field showing the current call’s consumption:<\/p>\n <pre><code>{\n \"usage\": {\n \"prompt_tokens\": 350, \/\/ Input tokens\n \"completion_tokens\": 820, \/\/ Output tokens\n \"total_tokens\": 1170 \/\/ Total\n }\n}<\/code><\/pre>\n\n <h3>6.2 Pre-Estimate with Tokenizer<\/h3>\n <pre><code>from transformers import AutoTokenizer\n\n# Load Qwen's tokenizer\ntokenizer = AutoTokenizer.from_pretrained(\n \"Qwen\/Qwen2.5-7B-Instruct\", trust_remote_code=True\n)\n\n# Estimate input token count\ntext = \"Explain what a token is in simple terms\"\ntokens = tokenizer.encode(text)\nprint(f\"Input token count: {len(tokens)}\")\n# Output: ~15-20<\/code><\/pre>\n\n <h3>6.3 Cost Calculation Function<\/h3>\n <pre><code>def calc_cost(input_tokens, output_tokens,\n input_price_per_million=0.8,\n output_price_per_million=4.8):\n \"\"\"Calculate single-call cost for Qwen 3.5-Plus\"\"\"\n input_cost = (input_tokens \/ 1_000_000) * input_price_per_million\n output_cost = (output_tokens \/ 1_000_000) * output_price_per_million\n return round(input_cost + output_cost, 6)\n\n# Example\ncost = calc_cost(500, 1200)\nprint(f\"Cost: \u00a5{cost}\") # Cost: \u00a50.00616<\/code><\/pre>\n<\/div>\n\n\n<div class=\"card\">\n <h2>7. Best Practices for Saving Tokens<\/h2>\n <ol>\n <li><strong>Streamline system prompts<\/strong>: Remove unnecessary context; every extra token is consumed in every call<\/li>\n <li><strong>Control conversation rounds<\/strong>: Long conversation history accumulation is a hidden cost driver; reset conversations periodically<\/li>\n <li><strong>Set reasonable max_tokens<\/strong>: Don’t default to 4096 or 8192; set based on actual response length needs<\/li>\n <li><strong>Downgrade model for simple tasks<\/strong>: Use Qwen-Long (\u00a50.5\/M tokens) for daily Q&A, Qwen-Max only for complex reasoning<\/li>\n <li><strong>Use Batch mode for bulk tasks<\/strong>: Processing large data offline? Batch mode saves 50%<\/li>\n <li><strong>Enable streaming<\/strong>: Users can stop mid-generation, avoiding unnecessary lengthy responses<\/li>\n <li><strong>Cache high-frequency Q&A<\/strong>: Return cached results directly for identical questions \u2014 zero token cost<\/li>\n <li><strong>Use free quotas<\/strong>: Activate the Bailian Platform to claim 70M+ free tokens for trial<\/li>\n <\/ol>\n<\/div>\n\n<div class=\"summary-box\">\n <h2>Key Takeaways<\/h2>\n <div class=\"s-row\"><div class=\"s-label\">Token Basics<\/div><div class=\"s-desc\">The smallest unit of text processing in LLMs; ~1-2 Chinese characters per token, ~1 English word per token<\/div><\/div>\n <div class=\"s-row\"><div class=\"s-label\">Consumption Formula<\/div><div class=\"s-desc\">Total = Input Tokens + Output Tokens; Cost = Input \u00d7 Input Price + Output \u00d7 Output Price<\/div><\/div>\n <div class=\"s-row\"><div class=\"s-label\">Input Sources<\/div><div class=\"s-desc\">System prompt + user message + conversation history + tool\/file context<\/div><\/div>\n <div class=\"s-row\"><div class=\"s-label\">Output Characteristics<\/div><div class=\"s-desc\">Usually 3-10\u00d7 more expensive than input; thinking mode adds extra reasoning tokens<\/div><\/div>\n <div class=\"s-row\"><div class=\"s-label\">Qwen Pricing<\/div><div class=\"s-desc\">Long: \u00a50.5\/M input; Plus: \u00a50.8\/M input; Max: \u00a52.5\/M input (China mainland)<\/div><\/div>\n <div class=\"s-row\"><div class=\"s-label\">Saving Core<\/div><div class=\"s-desc\">Streamline prompts + control rounds + choose the right model + Batch 50% off + use free quotas<\/div><\/div>\n<\/div>\n\n<\/div>\n\n<footer>\n <p>Based on Alibaba Cloud Tongyi (Qwen) official documentation and LLM token pricing principles<\/p>\n<\/footer>\n\n<\/div>\n\n<script>\nfunction setLang(lang) {\n document.querySelectorAll('.lang-section').forEach(el => el.classList.remove('active'));\n document.getElementById('lang-' + lang).classList.add('active');\n document.querySelectorAll('.lang-switch button').forEach(btn => btn.classList.remove('active'));\n document.getElementById('btn-' + lang).classList.add('active');\n document.documentElement.lang = lang === 'zh' ? 'zh-CN' : 'en';\n}\nfunction toggleTheme() {\n var html = document.documentElement;\n if (html.hasAttribute('data-theme')) {\n html.removeAttribute('data-theme');\n localStorage.removeItem('theme');\n } else {\n html.setAttribute('data-theme', 'dark');\n localStorage.setItem('theme', 'dark');\n }\n}\nif (localStorage.getItem('theme') === 'dark') {\n document.documentElement.setAttribute('data-theme', 'dark');\n}\n<\/script>\n\n<\/body>\n<\/html>\n","protected":false},"excerpt":{"rendered":"<p>Token \u5b8c\u5168\u79d1\u666e\uff1aLLM \u8c03\u7528\u6d88\u8017\u4e0e\u8ba1\u8d39\u6307\u5357 Token Guide \u4e2d\u6587 English ☼ […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[11,9],"tags":[],"class_list":["post-308","post","type-post","status-publish","format-standard","hentry","category-cad","category-ai"],"blocksy_meta":[],"_links":{"self":[{"href":"https:\/\/numsimlab.com\/index.php?rest_route=\/wp\/v2\/posts\/308","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/numsimlab.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/numsimlab.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/numsimlab.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/numsimlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=308"}],"version-history":[{"count":1,"href":"https:\/\/numsimlab.com\/index.php?rest_route=\/wp\/v2\/posts\/308\/revisions"}],"predecessor-version":[{"id":309,"href":"https:\/\/numsimlab.com\/index.php?rest_route=\/wp\/v2\/posts\/308\/revisions\/309"}],"wp:attachment":[{"href":"https:\/\/numsimlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=308"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/numsimlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=308"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/numsimlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=308"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}

{"id":308,"date":"2026-07-06T09:18:34","date_gmt":"2026-07-06T01:18:34","guid":{"rendered":"https:\/\/numsimlab.com\/?p=308"},"modified":"2026-07-06T09:18:34","modified_gmt":"2026-07-06T01:18:34","slug":"token-%e5%ae%8c%e5%85%a8%e7%a7%91%e6%99%ae%ef%bc%9allm-%e8%b0%83%e7%94%a8%e7%9a%84%e6%b6%88%e8%80%97%e4%b8%8e%e8%ae%a1%e8%b4%b9","status":"publish","type":"post","link":"https:\/\/numsimlab.com\/?p=308","title":{"rendered":"Token \u5b8c\u5168\u79d1\u666e\uff1aLLM \u8c03\u7528\u7684\u6d88\u8017\u4e0e\u8ba1\u8d39"},"content":{"rendered":"\n\n\n\n\n\n\nToken \u5b8c\u5168\u79d1\u666e\uff1aLLM \u8c03\u7528\u6d88\u8017\u4e0e\u8ba1\u8d39\u6307\u5357<\/title>\n<style>\n:root {\n --bg: #f8f9fb; --surface: #ffffff; --ink: #1a1a2e; --muted: #5a6578;\n --rule: #e2e5ea; --accent: #2563eb; --accent2: #059669; --accent3: #8b5cf6;\n --code-bg: #1e293b; --code-text: #e2e8f0;\n --warn-bg: #fffbeb; --warn-border: #f59e0b;\n --err-bg: #fef2f2; --err-border: #ef4444;\n --ok-bg: #ecfdf5; --ok-border: #10b981;\n --info-bg: #eff6ff; --info-border: #3b82f6;\n --tip-bg: #f5f3ff; --tip-border: #8b5cf6;\n}\n[data-theme=\"dark\"] {\n --bg: #0f172a; --surface: #1e293b; --ink: #f1f5f9; --muted: #94a3b8;\n --rule: #334155; --accent: #60a5fa; --accent2: #34d399; --accent3: #a78bfa;\n --code-bg: #020617; --code-text: #cbd5e1;\n --warn-bg: #2a1b0f; --warn-border: #f59e0b;\n --err-bg: #2a0f0f; --err-border: #ef4444;\n --ok-bg: #0f2a1b; --ok-border: #10b981;\n --info-bg: #172554; --info-border: #3b82f6;\n --tip-bg: #1a1530; --tip-border: #8b5cf6;\n}\n*{margin:0;padding:0;box-sizing:border-box}\nbody{font-family:-apple-system,BlinkMacSystemFont,\"Segoe UI\",Roboto,\"Helvetica Neue\",\"Noto Sans SC\",\"PingFang SC\",\"Microsoft YaHei\",sans-serif;background:var(--bg);color:var(--ink);line-height:1.75;font-size:15px;transition:background .3s,color .3s}\n.container{max-width:920px;margin:0 auto;padding:5rem 1.25rem 3rem}\n\n\/* Header *\/\nheader{position:fixed;top:0;left:0;right:0;z-index:100;background:var(--surface);border-bottom:1px solid var(--rule);padding:.85rem 0;transition:background .3s}\n.header-inner{max-width:920px;margin:0 auto;padding:0 1.25rem;display:flex;justify-content:space-between;align-items:center}\n.logo{font-size:1.15rem;font-weight:700;color:var(--accent);display:flex;align-items:center;gap:.5rem}\n.logo svg{width:24px;height:24px}\n.controls{display:flex;gap:.6rem;align-items:center}\n.lang-switch{display:flex;background:var(--bg);border:1px solid var(--rule);border-radius:6px;overflow:hidden;transition:background .3s}\n.lang-switch button{border:none;background:transparent;color:var(--muted);padding:.35rem .75rem;cursor:pointer;font-size:.82rem;transition:all .2s}\n.lang-switch button.active{background:var(--accent);color:#fff}\n.theme-toggle{background:var(--bg);border:1px solid var(--rule);border-radius:6px;padding:.35rem .6rem;cursor:pointer;color:var(--muted);font-size:1rem;line-height:1;transition:all .2s}\n\n\/* Page *\/\n.page-header{margin-bottom:2.5rem;text-align:center}\n.page-header h1{font-size:1.8rem;font-weight:800;letter-spacing:-0.02em;margin-bottom:.5rem}\n.page-header p{color:var(--muted);font-size:1rem;max-width:640px;margin:0 auto}\n.badge{display:inline-block;background:var(--accent3);color:#fff;padding:.25rem .75rem;border-radius:999px;font-size:.75rem;font-weight:600;margin-bottom:.8rem;text-transform:uppercase;letter-spacing:.05em}\n\n.card{background:var(--surface);border:1px solid var(--rule);border-radius:12px;padding:1.75rem 2rem;margin-bottom:1.5rem;transition:background .3s}\n.card h2{font-size:1.25rem;font-weight:700;margin-bottom:1rem;padding-bottom:.5rem;border-bottom:2px solid var(--accent);display:inline-block}\n.card h3{font-size:1.05rem;font-weight:600;margin:1.5rem 0 .5rem;color:var(--accent)}\n.card h3:first-child{margin-top:0}\n.card p,.card li{color:var(--muted);margin-bottom:.6rem}\n.card ul,.card ol{padding-left:1.5rem;margin-bottom:.75rem}\nli{margin-bottom:.4rem}\n\npre{background:var(--code-bg);color:var(--code-text);padding:1.1rem 1.25rem;border-radius:8px;overflow-x:auto;font-family:\"SF Mono\",Monaco,\"Cascadia Code\",Consolas,monospace;font-size:.85rem;line-height:1.6;margin:.75rem 0}\ncode{background:var(--bg);color:var(--accent);padding:.15rem .4rem;border-radius:4px;font-family:\"SF Mono\",Monaco,\"Cascadia Code\",Consolas,monospace;font-size:.87em}\npre code{background:transparent;color:inherit;padding:0}\n.cm{color:#94a3b8}\n\n.cmd{display:flex;align-items:center;gap:.75rem;background:var(--code-bg);color:var(--code-text);padding:.8rem 1.2rem;border-radius:8px;font-family:\"SF Mono\",Monaco,\"Cascadia Code\",Consolas,monospace;font-size:.87rem;margin:.6rem 0;overflow-x:auto}\n.cmd .p{color:#6ee7b7;user-select:none;flex-shrink:0}\n.cmd .c{white-space:nowrap}\n\n.callout{padding:1rem 1.25rem;border-radius:8px;margin:1rem 0;border-left:4px solid}\n.callout.warn{background:var(--warn-bg);border-color:var(--warn-border)}\n.callout.err{background:var(--err-bg);border-color:var(--err-border)}\n.callout.ok{background:var(--ok-bg);border-color:var(--ok-border)}\n.callout.info{background:var(--info-bg);border-color:var(--info-border)}\n.callout.tip{background:var(--tip-bg);border-color:var(--tip-border)}\n.callout strong{display:block;margin-bottom:.3rem;color:var(--ink)}\n\n.table-wrap{overflow-x:auto;margin:1rem 0}\ntable{width:100%;border-collapse:collapse;font-size:.89rem;min-width:500px}\nth,td{padding:.6rem .85rem;text-align:left;border-bottom:1px solid var(--rule)}\nth{background:var(--bg);font-weight:600;color:var(--ink);white-space:nowrap}\ntr:hover td{background:var(--bg)}\n\n.tag{display:inline-block;padding:.15rem .5rem;border-radius:4px;font-size:.75rem;font-weight:600;margin-right:.2rem}\n.tag-r{background:var(--accent);color:#fff}\n.tag-g{background:var(--accent2);color:#fff}\n.tag-o{background:var(--warn-bg);color:#d97706;border:1px solid var(--warn-border)}\n\n.compare-grid{display:grid;grid-template-columns:1fr 1fr;gap:1rem;margin:1rem 0}\n.compare-box{border:1px solid var(--rule);border-radius:8px;padding:1.1rem;background:var(--surface)}\n.compare-box .box-title{font-weight:700;font-size:.88rem;margin-bottom:.5rem;color:var(--accent)}\n\n.summary-box{background:var(--surface);border:2px solid var(--accent);border-radius:12px;padding:1.5rem 1.75rem;margin-top:1.5rem}\n.summary-box h2{border-bottom-color:var(--accent2);margin-top:0}\n.s-row{display:flex;gap:1.5rem;margin:.6rem 0;align-items:flex-start}\n.s-label{flex-shrink:0;font-weight:700;font-size:.9rem;color:var(--accent);min-width:120px}\n.s-desc{color:var(--muted);font-size:.92rem}\n\n.lang-section{display:none}\n.lang-section.active{display:block}\n\nfooter{text-align:center;color:var(--muted);font-size:.8rem;margin-top:2rem;padding-top:1rem;border-top:1px solid var(--rule)}\n\n@media(max-width:640px){\n .compare-grid{grid-template-columns:1fr}\n .page-header h1{font-size:1.5rem}\n .card{padding:1.25rem 1.5rem}\n .s-row{flex-direction:column;gap:.3rem}\n .header-inner{flex-wrap:wrap;gap:.5rem}\n}\n<\/style>\n<\/head>\n<body>\n\n<header>\n <div class=\"header-inner\">\n <div class=\"logo\">\n <svg viewBox=\"0 0 24 24\" fill=\"none\" stroke=\"currentColor\" stroke-width=\"2\" stroke-linecap=\"round\" stroke-linejoin=\"round\">\n <circle cx=\"12\" cy=\"12\" r=\"10\"\/><path d=\"M12 6v6l4 2\"\/>\n <\/svg>\n Token Guide\n <\/div>\n <div class=\"controls\">\n <div class=\"lang-switch\">\n <button class=\"active\" onclick=\"setLang('zh')\" id=\"btn-zh\">\u4e2d\u6587<\/button>\n <button onclick=\"setLang('en')\" id=\"btn-en\">English<\/button>\n <\/div>\n <button class=\"theme-toggle\" onclick=\"toggleTheme()\" title=\"Toggle theme\">☼<\/button>\n <\/div>\n <\/div>\n<\/header>\n\n<div class=\"container\">\n\n\n<div class=\"lang-section active\" id=\"lang-zh\">\n\n<div class=\"page-header\">\n <span class=\"badge\">Token \u79d1\u666e<\/span>\n <h1>Token \u5b8c\u5168\u79d1\u666e\uff1aLLM \u8c03\u7528\u7684\u6d88\u8017\u4e0e\u8ba1\u8d39<\/h1>\n <p>\u4ece\u96f6\u7406\u89e3 Token \u662f\u4ec0\u4e48\u3001\u5927\u6a21\u578b\u8c03\u7528\u8fc7\u7a0b\u4e2d Token \u5982\u4f55\u6d88\u8017\u3001\u4ee5\u53ca\u4ee5\u5343\u95ee\uff08Qwen\uff09\u7cfb\u5217\u4e3a\u4f8b\u7684\u8ba1\u8d39\u5168\u8c8c<\/p>\n<\/div>\n\n\n<div class=\"card\">\n <h2>\u4e00\u3001Token \u662f\u4ec0\u4e48\uff1f<\/h2>\n <p><strong>Token<\/strong> \u662f\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u5904\u7406\u6587\u672c\u7684<strong>\u6700\u5c0f\u57fa\u672c\u5355\u4f4d<\/strong>\u3002\u53ef\u4ee5\u628a\u5b83\u7406\u89e3\u4e3a\u6a21\u578b”\u9605\u8bfb”\u548c”\u5199\u4f5c”\u65f6\u4f7f\u7528\u7684”\u539f\u5b50\u8bcd\u6c47”\u2014\u2014\u6a21\u578b\u4e0d\u4f1a\u76f4\u63a5\u5904\u7406\u6c49\u5b57\u6216\u82f1\u6587\u5b57\u6bcd\uff0c\u800c\u662f\u5c06\u5b83\u4eec\u62c6\u89e3\u4e3a Token \u5e8f\u5217\u3002<\/p>\n\n <h3>Token \u5207\u5206\u793a\u4f8b<\/h3>\n <div class=\"table-wrap\">\n <table>\n <tr><th>\u539f\u59cb\u6587\u672c<\/th><th>Token \u6570\u91cf<\/th><th>Token \u5207\u5206\u7ed3\u679c<\/th><\/tr>\n <tr><td><code>\u4eba\u5de5\u667a\u80fd<\/code>\uff08\u4e2d\u6587\uff09<\/td><td>\u7ea6 2-4 \u4e2a<\/td><td><code>\u4eba\u5de5<\/code> + <code>\u667a\u80fd<\/code><\/td><\/tr>\n <tr><td><code>Hello World<\/code>\uff08\u82f1\u6587\uff09<\/td><td>\u7ea6 2-3 \u4e2a<\/td><td><code>Hello<\/code> + <code> World<\/code><\/td><\/tr>\n <tr><td><code>transformer<\/code>\uff08\u82f1\u6587\u957f\u8bcd\uff09<\/td><td>\u7ea6 2-3 \u4e2a<\/td><td><code>trans<\/code> + <code>former<\/code><\/td><\/tr>\n <tr><td><code>123456<\/code>\uff08\u6570\u5b57\uff09<\/td><td>1-3 \u4e2a<\/td><td><code>123<\/code> + <code>456<\/code><\/td><\/tr>\n <tr><td><code>print(\"hello\")<\/code>\uff08\u4ee3\u7801\uff09<\/td><td>\u7ea6 5-6 \u4e2a<\/td><td><code>print<\/code> + <code>(<\/code> + <code>\"<\/code> + <code>hello<\/code> + <code>\"<\/code> + <code>)<\/code><\/td><\/tr>\n <\/table>\n <\/div>\n\n <div class=\"callout info\">\n <strong>\u7ecf\u9a8c\u6cd5\u5219<\/strong>\n 1 \u4e2a\u4e2d\u6587\u5b57\u7b26 ≈ 1.5–2 \u4e2a Token | 1 \u4e2a\u82f1\u6587\u5355\u8bcd ≈ 1–1.5 \u4e2a Token | 1000 \u4e2a Token ≈ \u7ea6 750 \u4e2a\u82f1\u6587\u5355\u8bcd ≈ \u7ea6 500 \u4e2a\u4e2d\u6587\u6c49\u5b57\n <\/div>\n\n <h3>\u4e3a\u4ec0\u4e48\u9700\u8981 Token\uff1f<\/h3>\n <ul>\n <li><strong>\u6a21\u578b\u53ea\u80fd\u5904\u7406\u6570\u5b57<\/strong>\uff1aLLM \u5185\u90e8\u662f\u795e\u7ecf\u7f51\u7edc\uff0c\u6240\u6709\u8f93\u5165\u5fc5\u987b\u5148\u8f6c\u5316\u4e3a\u6570\u503c\u5411\u91cf\uff0cToken \u5c31\u662f\u6587\u672c\u5230\u6570\u5b57\u7684\u6865\u6881<\/li>\n <li><strong>\u8ba1\u8d39\u7684\u57fa\u7840\u5355\u5143<\/strong>\uff1aAPI \u6309 Token \u6570\u91cf\u6536\u8d39\uff0c\u7528\u591a\u5c11\u4ed8\u591a\u5c11\uff0c\u516c\u5e73\u900f\u660e<\/li>\n <li><strong>\u4e0a\u4e0b\u6587\u7a97\u53e3\u7684\u5ea6\u91cf<\/strong>\uff1a\u6a21\u578b\u80fd\u4e00\u6b21”\u8bb0\u4f4f”\u7684\u5185\u5bb9\u4e0a\u9650\u4ee5 Token \u6570\u8861\u91cf\uff08\u5982 128K Token\uff09<\/li>\n <\/ul>\n<\/div>\n\n\n<div class=\"card\">\n <h2>\u4e8c\u3001LLM \u8c03\u7528\u8fc7\u7a0b\u4e2d Token \u5982\u4f55\u6d88\u8017\uff1f<\/h2>\n <p>\u4e00\u6b21\u5b8c\u6574\u7684 LLM \u8c03\u7528\uff0cToken \u6d88\u8017\u53d1\u751f\u5728\u4e09\u4e2a\u73af\u8282\uff1a<\/p>\n\n <h3>2.1 \u7cfb\u7edf\u63d0\u793a\u8bcd\uff08System Prompt\uff09<\/h3>\n <p>\u5f00\u53d1\u8005\u9884\u8bbe\u7684”\u89d2\u8272\u8bbe\u5b9a”\uff0c\u5982”\u4f60\u662f\u4e00\u4e2a\u4e13\u4e1a\u7684\u6cd5\u5f8b\u987e\u95ee”\u3002\u6bcf\u6b21\u8c03\u7528\u90fd\u4f1a<strong>\u91cd\u590d\u8ba1\u5165<\/strong>\u8f93\u5165 Token\u3002\u7cfb\u7edf\u63d0\u793a\u8bcd\u8d8a\u957f\uff0c\u6bcf\u6b21\u5bf9\u8bdd\u7684\u56fa\u5b9a\u6210\u672c\u8d8a\u9ad8\u3002<\/p>\n <pre><code>system_prompt = \"\u4f60\u662f\u4e00\u4e2a\u5730\u8d28\u5efa\u6a21\u4e13\u5bb6\uff0c\u8bf7\u7528\u4e13\u4e1a\u672f\u8bed\u56de\u7b54\u3002\"\n# \u8fd9\u6761\u63d0\u793a\u8bcd\u7ea6 10-15 \u4e2a Token\uff0c\u6bcf\u6b21\u5bf9\u8bdd\u90fd\u4f1a\u6d88\u8017<\/code><\/pre>\n\n <h3>2.2 \u8f93\u5165 Token\uff08Input \/ Prompt Tokens\uff09<\/h3>\n <p>\u7528\u6237\u63d0\u51fa\u7684\u95ee\u9898\u6216\u7ed9\u6a21\u578b\u7684\u6240\u6709\u6587\u672c\u5185\u5bb9\uff0c\u5305\u62ec\uff1a<\/p>\n <ul>\n <li>\u7528\u6237\u6d88\u606f\u6b63\u6587<\/li>\n <li>\u5bf9\u8bdd\u5386\u53f2\uff08\u591a\u8f6e\u5bf9\u8bdd\u65f6\uff0c\u4e4b\u524d\u7684\u95ee\u7b54\u4f1a\u91cd\u65b0\u53d1\u9001\uff09<\/li>\n <li>\u4e0a\u4f20\u7684\u6587\u4ef6\u5185\u5bb9\uff08\u5982 PDF \u6458\u8981\u3001\u4ee3\u7801\u6587\u4ef6\uff09<\/li>\n <li>\u5de5\u5177\u8c03\u7528\u7ed3\u679c\uff08Function Calling \u7684\u8fd4\u56de\u503c\uff09<\/li>\n <\/ul>\n\n <div class=\"callout warn\">\n <strong>\u591a\u8f6e\u5bf9\u8bdd\u7684\u9690\u85cf\u6210\u672c<\/strong>\n \u6bcf\u6b21\u65b0\u5bf9\u8bdd\uff0c<strong>\u5168\u90e8\u5386\u53f2\u6d88\u606f<\/strong>\u4f1a\u88ab\u91cd\u65b0\u53d1\u9001\u3002\u7b2c 10 \u8f6e\u5bf9\u8bdd\u65f6\uff0c\u524d 9 \u8f6e\u7684\u95ee\u7b54\u5df2\u7d2f\u8ba1\u6210\u4e3a\u5927\u91cf\u8f93\u5165 Token\u3002\u8fd9\u5c31\u662f\u4e3a\u4ec0\u4e48\u957f\u5bf9\u8bdd\u8d8a\u6765\u8d8a\u8d35\u3002\n <\/div>\n\n <h3>2.3 \u8f93\u51fa Token\uff08Output \/ Completion Tokens\uff09<\/h3>\n <p>\u6a21\u578b\u751f\u6210\u7684\u56de\u590d\u6587\u672c\u3002\u8f93\u51fa Token \u6309\u5b9e\u9645\u751f\u6210\u6570\u91cf\u8ba1\u7b97\uff0c\u53d7 <code>max_tokens<\/code> \u53c2\u6570\u9650\u5236\u3002<\/p>\n\n <h3>2.4 \u5b8c\u6574\u6d88\u8017\u6d41\u7a0b<\/h3>\n <pre><code>\u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\n\u2502 \u4e00\u6b21 API \u8c03\u7528\u7684 Token \u8d26\u5355 \u2502\n\u251c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2524\n\u2502 \u8f93\u5165 Token\uff08\u4f60\u53d1\u7ed9\u6a21\u578b\u7684\uff09 \u2502\n\u2502 \u251c\u2500\u2500 \u7cfb\u7edf\u63d0\u793a\u8bcd\uff08System Prompt\uff09 \u2502\n\u2502 \u251c\u2500\u2500 \u7528\u6237\u6d88\u606f\uff08\u5f53\u524d\u95ee\u9898\uff09 \u2502\n\u2502 \u251c\u2500\u2500 \u5bf9\u8bdd\u5386\u53f2\uff08\u4e4b\u524d\u7684\u591a\u8f6e\u95ee\u7b54\uff09 \u2502\n\u2502 \u2514\u2500\u2500 \u5de5\u5177\/\u6587\u4ef6\u4e0a\u4e0b\u6587\uff08\u5982\u6709\uff09 \u2502\n\u251c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2524\n\u2502 \u8f93\u51fa Token\uff08\u6a21\u578b\u751f\u6210\u7684\u56de\u590d\uff09 \u2502\n\u2502 \u2514\u2500\u2500 \u6a21\u578b\u56de\u7b54\u7684\u5168\u90e8\u5185\u5bb9 \u2502\n\u251c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2524\n\u2502 \u603b\u6d88\u8017 = \u8f93\u5165 Token + \u8f93\u51fa Token \u2502\n\u2502 \u603b\u8d39\u7528 = \u8f93\u5165 Token \u00d7 \u8f93\u5165\u5355\u4ef7 + \u8f93\u51fa Token \u00d7 \u8f93\u51fa\u5355\u4ef7 \u2502\n\u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518<\/code><\/pre>\n<\/div>\n\n\n<div class=\"card\">\n <h2>\u4e09\u3001Token \u8ba1\u8d39\u7684\u57fa\u672c\u516c\u5f0f<\/h2>\n <div class=\"cmd\"><span class=\"p\">$<\/span><span class=\"c\">\u603b\u8d39\u7528 = (\u8f93\u5165Token\u6570 × \u8f93\u5165\u5355\u4ef7) + (\u8f93\u51faToken\u6570 × \u8f93\u51fa\u5355\u4ef7)<\/span><\/div>\n\n <h3>\u8ba1\u7b97\u793a\u4f8b<\/h3>\n <p>\u5047\u8bbe\u4f7f\u7528\u5343\u95ee Qwen-Long \u6a21\u578b\uff08\u8f93\u5165 0.5 \u5143\/\u767e\u4e07Token\uff0c\u8f93\u51fa 2 \u5143\/\u767e\u4e07Token\uff09\uff1a<\/p>\n <pre><code># \u7528\u6237\u63d0\u95ee\uff1a\"\u8bf7\u89e3\u91ca\u4ec0\u4e48\u662f Token\"\uff08\u7ea6 300 Token\uff09\n# \u6a21\u578b\u56de\u590d\u4e86\u8be6\u7ec6\u89e3\u91ca\uff08\u7ea6 900 Token\uff09\n\n\u8f93\u5165\u8d39\u7528 = 300 \/ 1,000,000 \u00d7 0.5 = \u00a50.00015\n\u8f93\u51fa\u8d39\u7528 = 900 \/ 1,000,000 \u00d7 2.0 = \u00a50.0018\n\u603b\u8d39\u7528 = \u00a50.00195\uff08\u7ea6 0.2 \u5206\u94b1\uff09\n\n# \u4e00\u5929\u8c03\u7528 1000 \u6b21 \u2248 \u00a51.95\n# \u4e00\u5929\u8c03\u7528 10 \u4e07\u6b21 \u2248 \u00a5195<\/code><\/pre>\n\n <div class=\"callout tip\">\n <strong>\u5173\u952e\u8ba4\u77e5<\/strong>\n \u8f93\u51fa Token \u7684\u5355\u4ef7\u901a\u5e38\u662f\u8f93\u5165 Token \u7684 <strong>3–10 \u500d<\/strong>\u3002\u56e0\u6b64\uff0c\u63a7\u5236\u6a21\u578b\u56de\u590d\u957f\u5ea6\uff08\u8bbe\u7f6e\u5408\u7406\u7684 <code>max_tokens<\/code>\uff09\u6bd4\u7cbe\u7b80\u8f93\u5165\u66f4\u7701\u94b1\u3002\n <\/div>\n\n <h3>\u591a\u8f6e\u5bf9\u8bdd\u7684 Token \u7d2f\u79ef<\/h3>\n <pre><code>\u7b2c 1 \u8f6e\uff1a\u8f93\u5165 500 Token \u2192 \u8f93\u51fa 800 Token \u2192 \u7d2f\u8ba1 1300\n\u7b2c 2 \u8f6e\uff1a\u8f93\u5165 500+1300=1800 Token \u2192 \u8f93\u51fa 600 Token \u2192 \u7d2f\u8ba1 3700\n\u7b2c 3 \u8f6e\uff1a\u8f93\u5165 500+3700=4200 Token \u2192 \u8f93\u51fa 500 Token \u2192 \u7d2f\u8ba1 8400\n...\n\u7b2c 10 \u8f6e\uff1a\u8f93\u5165\u53ef\u80fd\u5df2\u8d85\u8fc7 20000 Token\n\n# \u6bcf\u6b21\u5bf9\u8bdd\u589e\u957f\uff0c\u8f93\u5165 Token \u5448\u7ebf\u6027\u589e\u957f<\/code><\/pre>\n<\/div>\n\n\n<div class=\"card\">\n <h2>\u56db\u3001\u5343\u95ee\uff08Qwen\uff09\u7cfb\u5217\u8ba1\u8d39\u8be6\u89e3<\/h2>\n <p>\u963f\u91cc\u4e91\u5343\u95ee\u7cfb\u5217\u6309 Token \u8ba1\u8d39\uff0c\u4ef7\u683c\u56e0\u6a21\u578b\u3001\u8f93\u5165\u957f\u5ea6\u3001\u90e8\u7f72\u533a\u57df\u548c\u8c03\u7528\u6a21\u5f0f\u800c\u5f02\u3002\u4ee5\u4e0b\u4e3a 2026 \u5e74<strong>\u4e2d\u56fd\u5185\u5730\uff08\u5317\u4eac\u5730\u57df\uff09<\/strong>\u4e3b\u8981\u6a21\u578b\u4ef7\u683c\u53c2\u8003\uff1a<\/p>\n\n <div class=\"table-wrap\">\n <table>\n <thead>\n <tr><th>\u6a21\u578b<\/th><th>\u8f93\u5165\u5355\u4ef7\uff08\/\u767e\u4e07Token\uff09<\/th><th>\u8f93\u51fa\u5355\u4ef7\uff08\/\u767e\u4e07Token\uff09<\/th><th>\u7279\u70b9<\/th><\/tr>\n <\/thead>\n <tbody>\n <tr><td><strong>Qwen-Long<\/strong><\/td><td><strong>\u00a50.5<\/strong><\/td><td><strong>\u00a52.0<\/strong><\/td><td>\u957f\u6587\u672c\u4e13\u7528\uff0c\u6027\u4ef7\u6bd4\u6700\u9ad8<\/td><\/tr>\n <tr><td><strong>Qwen3.5-Plus<\/strong><\/td><td><strong>\u00a50.8<\/strong><\/td><td><strong>\u00a54.8<\/strong><\/td><td>\u4e3b\u529b\u6a21\u578b\uff0c\u591a\u6a21\u6001\uff0c0-128K \u4e0a\u4e0b\u6587<\/td><\/tr>\n <tr><td>Qwen3.5-Plus (128K-256K)<\/td><td>\u00a52.0<\/td><td>\u00a512.0<\/td><td>\u8d85\u957f\u8f93\u5165\u533a\u95f4<\/td><\/tr>\n <tr><td>Qwen3.5-Plus (256K-1M)<\/td><td>\u00a54.0<\/td><td>\u00a524.0<\/td><td>\u767e\u4e07\u7ea7\u4e0a\u4e0b\u6587\u533a\u95f4<\/td><\/tr>\n <tr><td><strong>Qwen3-Max<\/strong><\/td><td><strong>\u00a52.5<\/strong><\/td><td><strong>\u00a510.0<\/strong><\/td><td>\u65d7\u8230\u6a21\u578b\uff0c0-32K \u4e0a\u4e0b\u6587<\/td><\/tr>\n <tr><td>Qwen3-Max (32K-128K)<\/td><td>\u00a54.0<\/td><td>\u00a516.0<\/td><td>\u4e2d\u957f\u6587\u672c\u533a\u95f4<\/td><\/tr>\n <tr><td>Qwen3-Max (128K-252K)<\/td><td>\u00a57.0<\/td><td>\u00a528.0<\/td><td>\u8d85\u957f\u6587\u672c\u533a\u95f4<\/td><\/tr>\n <tr><td><strong>Qwen3-Omni-Flash<\/strong><\/td><td>\u00a51.8 <small>(\u6587\u672c)<\/small><\/td><td>\u00a56.9<\/td><td>\u591a\u6a21\u6001\uff08\u542b\u97f3\u9891\/\u56fe\u7247\/\u89c6\u9891\u8f93\u5165\uff09<\/td><\/tr>\n <tr><td>Qwen3-Omni-Flash <small>(\u97f3\u9891\u8f93\u5165)<\/small><\/td><td>\u00a515.8<\/td><td>\u00a512.7<\/td><td>\u97f3\u9891\u8f93\u5165\u4ef7\u683c\u66f4\u9ad8<\/td><\/tr>\n <\/tbody>\n <\/table>\n <\/div>\n\n <h3>\u7701\u94b1\u6280\u5de7<\/h3>\n <div class=\"compare-grid\">\n <div class=\"compare-box\">\n <div class=\"box-title\">Batch \u8c03\u7528\uff08\u534a\u4ef7\uff09<\/div>\n <p style=\"font-size:.85rem;\">\u4e0d\u8981\u6c42\u5b9e\u65f6\u54cd\u5e94\u7684\u4efb\u52a1\u53ef\u63d0\u4ea4\u6279\u91cf\u8bf7\u6c42\uff0c\u8f93\u5165\u548c\u8f93\u51fa\u5355\u4ef7\u5747\u4e3a\u5b9e\u65f6\u63a8\u7406\u7684 <strong>50%<\/strong>\u3002\u9002\u5408\u79bb\u7ebf\u5206\u6790\u3001\u6570\u636e\u6807\u6ce8\u7b49\u573a\u666f\u3002<\/p>\n <\/div>\n <div class=\"compare-box\">\n <div class=\"box-title\">\u4e0a\u4e0b\u6587\u7f13\u5b58<\/div>\n <p style=\"font-size:.85rem;\">\u9ad8\u9891\u91cd\u590d\u7684\u7cfb\u7edf\u63d0\u793a\u8bcd\u6216\u6587\u6863\u53ef\u7f13\u5b58\uff0c\u547d\u4e2d\u7f13\u5b58\u65f6<strong>\u4ec5\u8f93\u5165 Token \u4eab\u6298\u6263<\/strong>\u3002\u9002\u5408\u56fa\u5b9a Prompt \u7684\u591a\u6b21\u8c03\u7528\u3002<\/p>\n <\/div>\n <div class=\"compare-box\">\n <div class=\"box-title\">\u514d\u8d39\u989d\u5ea6<\/div>\n <p style=\"font-size:.85rem;\">\u65b0\u7528\u6237\u5f00\u901a\u963f\u91cc\u4e91\u767e\u70bc\u5e73\u53f0\u5373\u9001\u5404\u6a21\u578b <strong>100 \u4e07 Token<\/strong> \u514d\u8d39\u989d\u5ea6\uff0890 \u5929\u6709\u6548\uff09\uff0c\u8fd8\u53ef\u9886\u53d6\u8d85 7000 \u4e07 Token\u3002<\/p>\n <\/div>\n <div class=\"compare-box\">\n <div class=\"box-title\">\u5408\u7406\u9009\u62e9\u6a21\u578b<\/div>\n <p style=\"font-size:.85rem;\">\u7b80\u5355\u4efb\u52a1\u7528 Qwen-Long\uff08\u6700\u4f4e\u4ef7\uff09\uff0c\u590d\u6742\u63a8\u7406\u7528 Qwen-Max\u3002\u4e0d\u8981\u7528\u65d7\u8230\u6a21\u578b\u5904\u7406\u65e5\u5e38\u95f2\u804a\u3002<\/p>\n <\/div>\n <\/div>\n\n <div class=\"callout info\">\n <strong>\u4ef7\u683c\u53c2\u8003<\/strong>\n \u4ee5\u4e0a\u4ef7\u683c\u4e3a\u963f\u91cc\u4e91\u767e\u70bc\u5e73\u53f0\u4e2d\u56fd\u5185\u5730\uff08\u5317\u4eac\u5730\u57df\uff09\u516c\u5f00\u5b9a\u4ef7\uff0c\u5b9e\u9645\u4ef7\u683c\u4ee5\u5b98\u65b9\u6700\u65b0\u516c\u544a\u4e3a\u51c6\u3002\u56fd\u9645\u90e8\u7f72\uff08\u65b0\u52a0\u5761\uff09\u4ef7\u683c\u7ea6\u4e3a\u5185\u5730\u7684 3–5 \u500d\u3002<br>\n \u5b98\u65b9\u5b9a\u4ef7\u9875\uff1a<a href=\"https:\/\/www.aliyun.com\/product\/tongyi\" target=\"_blank\">aliyun.com\/product\/tongyi<\/a> | \u767e\u70bc\u5e73\u53f0\uff1a<a href=\"https:\/\/www.aliyun.com\/product\/bailian\" target=\"_blank\">aliyun.com\/product\/bailian<\/a>\n <\/div>\n<\/div>\n\n\n<div class=\"card\">\n <h2>\u4e94\u3001\u5f71\u54cd Token \u6d88\u8017\u7684\u989d\u5916\u56e0\u7d20<\/h2>\n <div class=\"table-wrap\">\n <table>\n <tr><th>\u56e0\u7d20<\/th><th>\u5f71\u54cd<\/th><th>\u8bf4\u660e<\/th><\/tr>\n <tr><td><strong>\u601d\u8003\u6a21\u5f0f\uff08Thinking\uff09<\/strong><\/td><td>\u8f93\u51fa Token \u5927\u5e45\u589e\u52a0<\/td><td>\u6a21\u578b\u5185\u90e8\u63a8\u7406\u8fc7\u7a0b\u4e5f\u8ba1\u5165\u8f93\u51fa Token\uff0c\u53ef\u80fd\u7ffb\u500d\u751a\u81f3\u66f4\u591a<\/td><\/tr>\n <tr><td><strong>Function Calling<\/strong><\/td><td>\u8f93\u5165 Token \u589e\u52a0<\/td><td>\u5de5\u5177\u5b9a\u4e49\u548c\u8fd4\u56de\u7ed3\u679c\u90fd\u8ba1\u5165\u8f93\u5165 Token<\/td><\/tr>\n <tr><td><strong>\u591a\u6a21\u6001\u8f93\u5165<\/strong><\/td><td>\u5355\u4ef7\u66f4\u9ad8<\/td><td>\u56fe\u7247\u3001\u97f3\u9891\u3001\u89c6\u9891\u6309\u5404\u81ea\u5355\u4ef7\u8ba1\u8d39\uff0c\u4e14\u901a\u5e38\u8fdc\u9ad8\u4e8e\u7eaf\u6587\u672c<\/td><\/tr>\n <tr><td><strong>max_tokens \u8bbe\u7f6e<\/strong><\/td><td>\u9650\u5236\u8f93\u51fa\u4e0a\u9650<\/td><td>\u8bbe\u7f6e\u8fc7\u5927\u4f1a\u5bfc\u81f4\u975e\u5fc5\u8981\u7684\u9ad8\u6d88\u8017\uff0c\u5efa\u8bae\u6839\u636e\u5b9e\u9645\u9700\u6c42\u8bbe\u5b9a\u5408\u7406\u503c<\/td><\/tr>\n <tr><td><strong>\u6d41\u5f0f\u8f93\u51fa\uff08Streaming\uff09<\/strong><\/td><td>\u4e0d\u51cf\u5c11\u603b\u6d88\u8017<\/td><td>\u4f46\u53ef\u4e2d\u9014\u505c\u6b62\u751f\u6210\uff0c\u907f\u514d\u4e0d\u5fc5\u8981\u8f93\u51fa<\/td><\/tr>\n <\/table>\n <\/div>\n\n <h3>\u601d\u8003\u6a21\u5f0f\u7684 Token \u5f00\u9500<\/h3>\n <pre><code># \u666e\u901a\u6a21\u5f0f\n\u8f93\u5165 500 Token \u2192 \u8f93\u51fa 800 Token \u2192 \u603b\u6d88\u8017 1300 Token\n\n# \u601d\u8003\u6a21\u5f0f\uff08\u5f00\u542f\u6df1\u5ea6\u63a8\u7406\uff09\n\u8f93\u5165 500 Token \u2192 \u601d\u8003\u8fc7\u7a0b 1200 Token + \u6700\u7ec8\u56de\u7b54 600 Token\n\u2192 \u8f93\u51fa 1800 Token \u2192 \u603b\u6d88\u8017 2300 Token\n# \u8d39\u7528\u53ef\u80fd\u662f\u666e\u901a\u6a21\u5f0f\u7684 2-3 \u500d<\/code><\/pre>\n<\/div>\n\n\n<div class=\"card\">\n <h2>\u516d\u3001\u5982\u4f55\u8ba1\u7b97\u548c\u76d1\u63a7 Token \u6d88\u8017<\/h2>\n <h3>6.1 API \u8fd4\u56de\u503c\u4e2d\u67e5\u770b<\/h3>\n <p>\u6bcf\u6b21\u8c03\u7528 API \u540e\uff0c\u54cd\u5e94\u4e2d\u4f1a\u5305\u542b <code>usage<\/code> \u5b57\u6bb5\uff0c\u76f4\u63a5\u663e\u793a\u672c\u6b21\u6d88\u8017\uff1a<\/p>\n <pre><code>{\n \"usage\": {\n \"prompt_tokens\": 350, \/\/ \u8f93\u5165 Token\n \"completion_tokens\": 820, \/\/ \u8f93\u51fa Token\n \"total_tokens\": 1170 \/\/ \u603b\u6d88\u8017\n }\n}<\/code><\/pre>\n\n <h3>6.2 \u4f7f\u7528 Tokenizer \u9884\u5148\u4f30\u7b97<\/h3>\n <pre><code>from transformers import AutoTokenizer\n\n# \u52a0\u8f7d\u5343\u95ee\u7684 Tokenizer\ntokenizer = AutoTokenizer.from_pretrained(\n \"Qwen\/Qwen2.5-7B-Instruct\", trust_remote_code=True\n)\n\n# \u4f30\u7b97\u8f93\u5165 Token \u6570\ntext = \"\u8bf7\u7528\u901a\u4fd7\u7684\u8bed\u8a00\u89e3\u91ca\u4ec0\u4e48\u662f Token\"\ntokens = tokenizer.encode(text)\nprint(f\"\u8f93\u5165 Token \u6570: {len(tokens)}\")\n# \u8f93\u51fa: \u8f93\u5165 Token \u6570: ~15-20<\/code><\/pre>\n\n <h3>6.3 \u6210\u672c\u8ba1\u7b97\u51fd\u6570<\/h3>\n <pre><code>def calc_cost(input_tokens, output_tokens,\n input_price_per_million=0.8,\n output_price_per_million=4.8):\n \"\"\"\u8ba1\u7b97\u5343\u95ee 3.5-Plus \u7684\u5355\u6b21\u8c03\u7528\u8d39\u7528\"\"\"\n input_cost = (input_tokens \/ 1_000_000) * input_price_per_million\n output_cost = (output_tokens \/ 1_000_000) * output_price_per_million\n return round(input_cost + output_cost, 6)\n\n# \u793a\u4f8b\ncost = calc_cost(500, 1200)\nprint(f\"\u8d39\u7528: \u00a5{cost}\") # \u8d39\u7528: \u00a50.00616<\/code><\/pre>\n<\/div>\n\n\n<div class=\"card\">\n <h2>\u4e03\u3001\u7701\u94b1\u6700\u4f73\u5b9e\u8df5<\/h2>\n <ol>\n <li><strong>\u7cbe\u7b80\u7cfb\u7edf\u63d0\u793a\u8bcd<\/strong>\uff1a\u53bb\u6389\u4e0d\u5fc5\u8981\u7684\u80cc\u666f\u8bf4\u660e\uff0c\u6bcf\u6761\u591a\u51fa\u7684 Token \u90fd\u4f1a\u5728\u6bcf\u6b21\u5bf9\u8bdd\u4e2d\u91cd\u590d\u6d88\u8017<\/li>\n <li><strong>\u63a7\u5236\u5bf9\u8bdd\u8f6e\u6b21<\/strong>\uff1a\u957f\u5bf9\u8bdd\u7684\u5386\u53f2\u7d2f\u79ef\u662f\u9690\u85cf\u7684\u6210\u672c\u5927\u5934\uff0c\u5fc5\u8981\u65f6\u53ef\u5b9a\u671f”\u91cd\u7f6e\u5bf9\u8bdd”<\/li>\n <li><strong>\u8bbe\u7f6e\u5408\u7406\u7684 max_tokens<\/strong>\uff1a\u4e0d\u8981\u9ed8\u8ba4\u8bbe 4096 \u6216 8192\uff0c\u6309\u5b9e\u9645\u56de\u590d\u957f\u5ea6\u9700\u6c42\u8bbe\u5b9a<\/li>\n <li><strong>\u7b80\u5355\u4efb\u52a1\u964d\u7ea7\u6a21\u578b<\/strong>\uff1a\u65e5\u5e38\u95ee\u7b54\u7528 Qwen-Long\uff080.5 \u5143\/\u767e\u4e07Token\uff09\uff0c\u590d\u6742\u63a8\u7406\u624d\u7528 Qwen-Max<\/li>\n <li><strong>\u6279\u91cf\u4efb\u52a1\u7528 Batch \u6a21\u5f0f<\/strong>\uff1a\u79bb\u7ebf\u5904\u7406\u5927\u6279\u91cf\u6570\u636e\u65f6\uff0cBatch \u6a21\u5f0f\u53ef\u8282\u7701 50% \u8d39\u7528<\/li>\n <li><strong>\u542f\u7528\u6d41\u5f0f\u8f93\u51fa<\/strong>\uff1a\u7528\u6237\u53ef\u4e2d\u9014\u505c\u6b62\u751f\u6210\uff0c\u907f\u514d\u6a21\u578b\u4ea7\u51fa\u4e0d\u9700\u8981\u7684\u5197\u957f\u56de\u590d<\/li>\n <li><strong>\u9ad8\u9891\u95ee\u7b54\u505a\u7f13\u5b58<\/strong>\uff1a\u76f8\u540c\u95ee\u9898\u76f4\u63a5\u8fd4\u56de\u7f13\u5b58\u7ed3\u679c\uff0c\u96f6 Token \u6d88\u8017<\/li>\n <li><strong>\u5229\u7528\u514d\u8d39\u989d\u5ea6<\/strong>\uff1a\u5f00\u901a\u767e\u70bc\u5e73\u53f0\u9886\u53d6 7000 \u4e07+ Token \u514d\u8d39\u989d\u5ea6\uff0c\u5148\u8bd5\u7528\u518d\u4ed8\u8d39<\/li>\n <\/ol>\n<\/div>\n\n<div class=\"summary-box\">\n <h2>\u6838\u5fc3\u8981\u70b9\u603b\u7ed3<\/h2>\n <div class=\"s-row\"><div class=\"s-label\">Token \u672c\u8d28<\/div><div class=\"s-desc\">LLM \u5904\u7406\u6587\u672c\u7684\u6700\u5c0f\u5355\u4f4d\uff0c\u4e2d\u6587\u7ea6 1-2 \u5b57\/Token\uff0c\u82f1\u6587\u7ea6 1 \u8bcd\/Token<\/div><\/div>\n <div class=\"s-row\"><div class=\"s-label\">\u6d88\u8017\u516c\u5f0f<\/div><div class=\"s-desc\">\u603b\u6d88\u8017 = \u8f93\u5165 Token + \u8f93\u51fa Token\uff1b\u603b\u8d39\u7528 = \u8f93\u5165 \u00d7 \u8f93\u5165\u5355\u4ef7 + \u8f93\u51fa \u00d7 \u8f93\u51fa\u5355\u4ef7<\/div><\/div>\n <div class=\"s-row\"><div class=\"s-label\">\u8f93\u5165\u6765\u6e90<\/div><div class=\"s-desc\">\u7cfb\u7edf\u63d0\u793a\u8bcd + \u7528\u6237\u6d88\u606f + \u5bf9\u8bdd\u5386\u53f2 + \u5de5\u5177\/\u6587\u4ef6\u4e0a\u4e0b\u6587<\/div><\/div>\n <div class=\"s-row\"><div class=\"s-label\">\u8f93\u51fa\u7279\u70b9<\/div><div class=\"s-desc\">\u5355\u4ef7\u901a\u5e38\u662f\u8f93\u5165\u7684 3-10 \u500d\uff0c\u601d\u8003\u6a21\u5f0f\u989d\u5916\u589e\u52a0\u63a8\u7406 Token<\/div><\/div>\n <div class=\"s-row\"><div class=\"s-label\">\u5343\u95ee\u4ef7\u683c<\/div><div class=\"s-desc\">Long \u6700\u4f4e \u00a50.5\/\u767e\u4e07Token \u8f93\u5165\uff1bPlus \u4e3b\u529b \u00a50.8 \u8f93\u5165\uff1bMax \u65d7\u8230 \u00a52.5 \u8f93\u5165<\/div><\/div>\n <div class=\"s-row\"><div class=\"s-label\">\u7701\u94b1\u6838\u5fc3<\/div><div class=\"s-desc\">\u7cbe\u7b80 Prompt + \u63a7\u5236\u5bf9\u8bdd\u8f6e\u6b21 + \u9009\u5bf9\u6a21\u578b + Batch \u534a\u4ef7 + \u5229\u7528\u514d\u8d39\u989d\u5ea6<\/div><\/div>\n<\/div>\n\n<\/div>\n\n\n<div class=\"lang-section\" id=\"lang-en\">\n\n<div class=\"page-header\">\n <span class=\"badge\">Token Guide<\/span>\n <h1>Complete Token Guide: LLM Consumption & Pricing<\/h1>\n <p>Understand what tokens are, how they are consumed during LLM calls, and a complete pricing overview using the Qwen model series as an example<\/p>\n<\/div>\n\n\n<div class=\"card\">\n <h2>1. What Is a Token?<\/h2>\n <p>A <strong>Token<\/strong> is the <strong>smallest basic unit<\/strong> that a Large Language Model (LLM) uses to process text. Think of it as the “atomic vocabulary” the model uses to “read” and “write” \u2014 the model doesn’t process Chinese characters or English letters directly, but breaks them into token sequences.<\/p>\n\n <h3>Tokenization Examples<\/h3>\n <div class=\"table-wrap\">\n <table>\n <tr><th>Original Text<\/th><th>Token Count<\/th><th>Token Breakdown<\/th><\/tr>\n <tr><td><code>artificial intelligence<\/code><\/td><td>~2-3<\/td><td><code>art<\/code> + <code>ificial<\/code> + <code> intelligence<\/code><\/td><\/tr>\n <tr><td><code>Hello World<\/code><\/td><td>~2-3<\/td><td><code>Hello<\/code> + <code> World<\/code><\/td><\/tr>\n <tr><td><code>transformer<\/code><\/td><td>~2-3<\/td><td><code>trans<\/code> + <code>former<\/code><\/td><\/tr>\n <tr><td><code>123456<\/code><\/td><td>1-3<\/td><td><code>123<\/code> + <code>456<\/code><\/td><\/tr>\n <tr><td><code>print(\"hello\")<\/code><\/td><td>~5-6<\/td><td><code>print<\/code> + <code>(<\/code> + <code>\"<\/code> + <code>hello<\/code> + <code>\"<\/code> + <code>)<\/code><\/td><\/tr>\n <\/table>\n <\/div>\n\n <div class=\"callout info\">\n <strong>Rule of Thumb<\/strong>\n 1 Chinese character ≈ 1.5–2 tokens | 1 English word ≈ 1–1.5 tokens | 1000 tokens ≈ ~750 English words ≈ ~500 Chinese characters\n <\/div>\n\n <h3>Why Tokens?<\/h3>\n <ul>\n <li><strong>Models only process numbers<\/strong>: LLMs are neural networks; all input must be converted to numerical vectors first. Tokens bridge text to numbers.<\/li>\n <li><strong>Unit of billing<\/strong>: APIs charge by token count \u2014 pay for what you use, transparent and fair.<\/li>\n <li><strong>Context window metric<\/strong>: The maximum amount of content a model can “remember” at once is measured in tokens (e.g., 128K tokens).<\/li>\n <\/ul>\n<\/div>\n\n\n<div class=\"card\">\n <h2>2. How Tokens Are Consumed During an LLM Call<\/h2>\n <p>A complete LLM call consumes tokens across three stages:<\/p>\n\n <h3>2.1 System Prompt<\/h3>\n <p>The developer-defined “role setting,” such as “You are a professional legal advisor.” It is <strong>re-counted every call<\/strong> into input tokens. The longer the system prompt, the higher the fixed cost per conversation.<\/p>\n <pre><code>system_prompt = \"You are a geological modeling expert. Use professional terms.\"\n# This prompt is ~10-15 tokens, consumed every conversation<\/code><\/pre>\n\n <h3>2.2 Input Tokens (Prompt Tokens)<\/h3>\n <p>All text content sent to the model, including:<\/p>\n <ul>\n <li>User message body<\/li>\n <li>Conversation history (prior Q&A rounds are re-sent in multi-turn chats)<\/li>\n <li>Uploaded file content (e.g., PDF excerpts, code files)<\/li>\n <li>Function calling results (tool return values)<\/li>\n <\/ul>\n\n <div class=\"callout warn\">\n <strong>Hidden cost of multi-turn conversations<\/strong>\n Every new turn, <strong>all historical messages<\/strong> are re-sent. By round 10, the previous 9 rounds have accumulated into a massive input token count. This is why long conversations get progressively more expensive.\n <\/div>\n\n <h3>2.3 Output Tokens (Completion Tokens)<\/h3>\n <p>The text the model generates in response. Output tokens are counted based on actual generation, limited by the <code>max_tokens<\/code> parameter.<\/p>\n\n <h3>2.4 Complete Consumption Flow<\/h3>\n <pre><code>\u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\n\u2502 Token Bill for One API Call \u2502\n\u251c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2524\n\u2502 Input Tokens (what you send to the model) \u2502\n\u2502 \u251c\u2500\u2500 System Prompt \u2502\n\u2502 \u251c\u2500\u2500 User Message (current question) \u2502\n\u2502 \u251c\u2500\u2500 Conversation History (prior Q&A rounds) \u2502\n\u2502 \u2514\u2500\u2500 Tool\/File Context (if any) \u2502\n\u251c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2524\n\u2502 Output Tokens (model's response) \u2502\n\u2502 \u2514\u2500\u2500 Full content of model's answer \u2502\n\u251c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2524\n\u2502 Total = Input Tokens + Output Tokens \u2502\n\u2502 Cost = Input \u00d7 Input Price + Output \u00d7 Output Price\u2502\n\u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518<\/code><\/pre>\n<\/div>\n\n\n<div class=\"card\">\n <h2>3. Basic Token Pricing Formula<\/h2>\n <div class=\"cmd\"><span class=\"p\">$<\/span><span class=\"c\">Total Cost = (Input Tokens × Input Price) + (Output Tokens × Output Price)<\/span><\/div>\n\n <h3>Calculation Example<\/h3>\n <p>Using Qwen-Long (input \u00a50.5\/million tokens, output \u00a52.0\/million tokens):<\/p>\n <pre><code># User asks: \"Explain what a token is\" (~300 tokens)\n# Model responds with detailed explanation (~900 tokens)\n\nInput cost = 300 \/ 1,000,000 \u00d7 0.5 = \u00a50.00015\nOutput cost = 900 \/ 1,000,000 \u00d7 2.0 = \u00a50.0018\nTotal cost = \u00a50.00195 (~0.2 Chinese cents)\n\n# 1,000 calls\/day \u2248 \u00a51.95\n# 100,000 calls\/day \u2248 \u00a5195<\/code><\/pre>\n\n <div class=\"callout tip\">\n <strong>Key Insight<\/strong>\n Output token pricing is typically <strong>3–10\u00d7<\/strong> higher than input. So controlling response length (setting a reasonable <code>max_tokens<\/code>) saves more than trimming input.\n <\/div>\n\n <h3>Multi-Turn Token Accumulation<\/h3>\n <pre><code>Round 1: Input 500 \u2192 Output 800 \u2192 Cumulative 1300\nRound 2: Input 500+1300=1800 \u2192 Output 600 \u2192 Cumulative 3700\nRound 3: Input 500+3700=4200 \u2192 Output 500 \u2192 Cumulative 8400\n...\nRound 10: Input may exceed 20,000 tokens\n\n# Input tokens grow linearly with each turn<\/code><\/pre>\n<\/div>\n\n\n<div class=\"card\">\n <h2>4. Qwen Series Pricing Breakdown<\/h2>\n <p>Alibaba Cloud’s Qwen series charges by token, with prices varying by model, input length, deployment region, and call mode. Below is the pricing reference for <strong>China mainland (Beijing region)<\/strong> as of 2026:<\/p>\n\n <div class=\"table-wrap\">\n <table>\n <thead>\n <tr><th>Model<\/th><th>Input (per M tokens)<\/th><th>Output (per M tokens)<\/th><th>Features<\/th><\/tr>\n <\/thead>\n <tbody>\n <tr><td><strong>Qwen-Long<\/strong><\/td><td><strong>\u00a50.5<\/strong><\/td><td><strong>\u00a52.0<\/strong><\/td><td>Long-text specialist, best value<\/td><\/tr>\n <tr><td><strong>Qwen3.5-Plus<\/strong><\/td><td><strong>\u00a50.8<\/strong><\/td><td><strong>\u00a54.8<\/strong><\/td><td>Main model, multimodal, 0-128K context<\/td><\/tr>\n <tr><td>Qwen3.5-Plus (128K-256K)<\/td><td>\u00a52.0<\/td><td>\u00a512.0<\/td><td>Extended input range<\/td><\/tr>\n <tr><td>Qwen3.5-Plus (256K-1M)<\/td><td>\u00a54.0<\/td><td>\u00a524.0<\/td><td>Million-token context range<\/td><\/tr>\n <tr><td><strong>Qwen3-Max<\/strong><\/td><td><strong>\u00a52.5<\/strong><\/td><td><strong>\u00a510.0<\/strong><\/td><td>Flagship model, 0-32K context<\/td><\/tr>\n <tr><td>Qwen3-Max (32K-128K)<\/td><td>\u00a54.0<\/td><td>\u00a516.0<\/td><td>Mid-long text range<\/td><\/tr>\n <tr><td>Qwen3-Max (128K-252K)<\/td><td>\u00a57.0<\/td><td>\u00a528.0<\/td><td>Ultra-long text range<\/td><\/tr>\n <tr><td><strong>Qwen3-Omni-Flash<\/strong><\/td><td>\u00a51.8 <small>(text)<\/small><\/td><td>\u00a56.9<\/td><td>Multimodal (audio\/image\/video input)<\/td><\/tr>\n <tr><td>Qwen3-Omni-Flash <small>(audio input)<\/small><\/td><td>\u00a515.8<\/td><td>\u00a512.7<\/td><td>Higher audio input pricing<\/td><\/tr>\n <\/tbody>\n <\/table>\n <\/div>\n\n <h3>Cost-Saving Tips<\/h3>\n <div class=\"compare-grid\">\n <div class=\"compare-box\">\n <div class=\"box-title\">Batch Calling (50% off)<\/div>\n <p style=\"font-size:.85rem;\">For non-real-time tasks, submit batch requests. Both input and output pricing is <strong>50%<\/strong> of real-time inference. Suitable for offline analysis and data labeling.<\/p>\n <\/div>\n <div class=\"compare-box\">\n <div class=\"box-title\">Context Caching<\/div>\n <p style=\"font-size:.85rem;\">High-frequency system prompts or documents can be cached. When the cache is hit, <strong>input tokens get a discount<\/strong>. Ideal for repeated calls with fixed prompts.<\/p>\n <\/div>\n <div class=\"compare-box\">\n <div class=\"box-title\">Free Quota<\/div>\n <p style=\"font-size:.85rem;\">New users of Alibaba Cloud Bailian Platform receive <strong>1 million tokens<\/strong> per model (90 days) plus over 70 million additional free tokens.<\/p>\n <\/div>\n <div class=\"compare-box\">\n <div class=\"box-title\">Choose the Right Model<\/div>\n <p style=\"font-size:.85rem;\">Use Qwen-Long (lowest price) for simple tasks; Qwen-Max for complex reasoning. Don’t use the flagship model for casual chat.<\/p>\n <\/div>\n <\/div>\n\n <div class=\"callout info\">\n <strong>Pricing Reference<\/strong>\n Prices above are public pricing for Alibaba Cloud Bailian Platform (Beijing region). Actual prices are subject to the latest official announcements. International deployment (Singapore) is approximately 3–5\u00d7 higher.<br>\n Official pricing: <a href=\"https:\/\/www.aliyun.com\/product\/tongyi\" target=\"_blank\">aliyun.com\/product\/tongyi<\/a> | Bailian Platform: <a href=\"https:\/\/www.aliyun.com\/product\/bailian\" target=\"_blank\">aliyun.com\/product\/bailian<\/a>\n <\/div>\n<\/div>\n\n\n<div class=\"card\">\n <h2>5. Additional Factors Affecting Token Consumption<\/h2>\n <div class=\"table-wrap\">\n <table>\n <tr><th>Factor<\/th><th>Impact<\/th><th>Description<\/th><\/tr>\n <tr><td><strong>Thinking Mode<\/strong><\/td><td>Output tokens increase significantly<\/td><td>Internal reasoning process is also counted as output, potentially doubling or more<\/td><\/tr>\n <tr><td><strong>Function Calling<\/strong><\/td><td>Input tokens increase<\/td><td>Tool definitions and return values are counted as input tokens<\/td><\/tr>\n <tr><td><strong>Multimodal Input<\/strong><\/td><td>Higher unit price<\/td><td>Images, audio, and video are billed at their own rates, usually much higher than plain text<\/td><\/tr>\n <tr><td><strong>max_tokens setting<\/strong><\/td><td>Caps output<\/td><td>Setting it too high leads to unnecessary consumption; set a reasonable value based on actual needs<\/td><\/tr>\n <tr><td><strong>Streaming<\/strong><\/td><td>Doesn’t reduce total<\/td><td>But allows mid-generation stopping to avoid unnecessary output<\/td><\/tr>\n <\/table>\n <\/div>\n\n <h3>Thinking Mode Token Overhead<\/h3>\n <pre><code># Normal mode\nInput 500 tokens \u2192 Output 800 tokens \u2192 Total 1300 tokens\n\n# Thinking mode (deep reasoning enabled)\nInput 500 tokens \u2192 Reasoning 1200 tokens + Final answer 600 tokens\n\u2192 Output 1800 tokens \u2192 Total 2300 tokens\n# Cost may be 2-3\u00d7 of normal mode<\/code><\/pre>\n<\/div>\n\n\n<div class=\"card\">\n <h2>6. How to Calculate and Monitor Token Consumption<\/h2>\n <h3>6.1 Check API Response<\/h3>\n <p>Every API response includes a <code>usage<\/code> field showing the current call’s consumption:<\/p>\n <pre><code>{\n \"usage\": {\n \"prompt_tokens\": 350, \/\/ Input tokens\n \"completion_tokens\": 820, \/\/ Output tokens\n \"total_tokens\": 1170 \/\/ Total\n }\n}<\/code><\/pre>\n\n <h3>6.2 Pre-Estimate with Tokenizer<\/h3>\n <pre><code>from transformers import AutoTokenizer\n\n# Load Qwen's tokenizer\ntokenizer = AutoTokenizer.from_pretrained(\n \"Qwen\/Qwen2.5-7B-Instruct\", trust_remote_code=True\n)\n\n# Estimate input token count\ntext = \"Explain what a token is in simple terms\"\ntokens = tokenizer.encode(text)\nprint(f\"Input token count: {len(tokens)}\")\n# Output: ~15-20<\/code><\/pre>\n\n <h3>6.3 Cost Calculation Function<\/h3>\n <pre><code>def calc_cost(input_tokens, output_tokens,\n input_price_per_million=0.8,\n output_price_per_million=4.8):\n \"\"\"Calculate single-call cost for Qwen 3.5-Plus\"\"\"\n input_cost = (input_tokens \/ 1_000_000) * input_price_per_million\n output_cost = (output_tokens \/ 1_000_000) * output_price_per_million\n return round(input_cost + output_cost, 6)\n\n# Example\ncost = calc_cost(500, 1200)\nprint(f\"Cost: \u00a5{cost}\") # Cost: \u00a50.00616<\/code><\/pre>\n<\/div>\n\n\n<div class=\"card\">\n <h2>7. Best Practices for Saving Tokens<\/h2>\n <ol>\n <li><strong>Streamline system prompts<\/strong>: Remove unnecessary context; every extra token is consumed in every call<\/li>\n <li><strong>Control conversation rounds<\/strong>: Long conversation history accumulation is a hidden cost driver; reset conversations periodically<\/li>\n <li><strong>Set reasonable max_tokens<\/strong>: Don’t default to 4096 or 8192; set based on actual response length needs<\/li>\n <li><strong>Downgrade model for simple tasks<\/strong>: Use Qwen-Long (\u00a50.5\/M tokens) for daily Q&A, Qwen-Max only for complex reasoning<\/li>\n <li><strong>Use Batch mode for bulk tasks<\/strong>: Processing large data offline? Batch mode saves 50%<\/li>\n <li><strong>Enable streaming<\/strong>: Users can stop mid-generation, avoiding unnecessary lengthy responses<\/li>\n <li><strong>Cache high-frequency Q&A<\/strong>: Return cached results directly for identical questions \u2014 zero token cost<\/li>\n <li><strong>Use free quotas<\/strong>: Activate the Bailian Platform to claim 70M+ free tokens for trial<\/li>\n <\/ol>\n<\/div>\n\n<div class=\"summary-box\">\n <h2>Key Takeaways<\/h2>\n <div class=\"s-row\"><div class=\"s-label\">Token Basics<\/div><div class=\"s-desc\">The smallest unit of text processing in LLMs; ~1-2 Chinese characters per token, ~1 English word per token<\/div><\/div>\n <div class=\"s-row\"><div class=\"s-label\">Consumption Formula<\/div><div class=\"s-desc\">Total = Input Tokens + Output Tokens; Cost = Input \u00d7 Input Price + Output \u00d7 Output Price<\/div><\/div>\n <div class=\"s-row\"><div class=\"s-label\">Input Sources<\/div><div class=\"s-desc\">System prompt + user message + conversation history + tool\/file context<\/div><\/div>\n <div class=\"s-row\"><div class=\"s-label\">Output Characteristics<\/div><div class=\"s-desc\">Usually 3-10\u00d7 more expensive than input; thinking mode adds extra reasoning tokens<\/div><\/div>\n <div class=\"s-row\"><div class=\"s-label\">Qwen Pricing<\/div><div class=\"s-desc\">Long: \u00a50.5\/M input; Plus: \u00a50.8\/M input; Max: \u00a52.5\/M input (China mainland)<\/div><\/div>\n <div class=\"s-row\"><div class=\"s-label\">Saving Core<\/div><div class=\"s-desc\">Streamline prompts + control rounds + choose the right model + Batch 50% off + use free quotas<\/div><\/div>\n<\/div>\n\n<\/div>\n\n<footer>\n <p>Based on Alibaba Cloud Tongyi (Qwen) official documentation and LLM token pricing principles<\/p>\n<\/footer>\n\n<\/div>\n\n<script>\nfunction setLang(lang) {\n document.querySelectorAll('.lang-section').forEach(el => el.classList.remove('active'));\n document.getElementById('lang-' + lang).classList.add('active');\n document.querySelectorAll('.lang-switch button').forEach(btn => btn.classList.remove('active'));\n document.getElementById('btn-' + lang).classList.add('active');\n document.documentElement.lang = lang === 'zh' ? 'zh-CN' : 'en';\n}\nfunction toggleTheme() {\n var html = document.documentElement;\n if (html.hasAttribute('data-theme')) {\n html.removeAttribute('data-theme');\n localStorage.removeItem('theme');\n } else {\n html.setAttribute('data-theme', 'dark');\n localStorage.setItem('theme', 'dark');\n }\n}\nif (localStorage.getItem('theme') === 'dark') {\n document.documentElement.setAttribute('data-theme', 'dark');\n}\n<\/script>\n\n<\/body>\n<\/html>\n","protected":false},"excerpt":{"rendered":"<p>Token \u5b8c\u5168\u79d1\u666e\uff1aLLM \u8c03\u7528\u6d88\u8017\u4e0e\u8ba1\u8d39\u6307\u5357 Token Guide \u4e2d\u6587 English ☼ […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[11,9],"tags":[],"class_list":["post-308","post","type-post","status-publish","format-standard","hentry","category-cad","category-ai"],"blocksy_meta":[],"_links":{"self":[{"href":"https:\/\/numsimlab.com\/index.php?rest_route=\/wp\/v2\/posts\/308","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/numsimlab.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/numsimlab.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/numsimlab.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/numsimlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=308"}],"version-history":[{"count":1,"href":"https:\/\/numsimlab.com\/index.php?rest_route=\/wp\/v2\/posts\/308\/revisions"}],"predecessor-version":[{"id":309,"href":"https:\/\/numsimlab.com\/index.php?rest_route=\/wp\/v2\/posts\/308\/revisions\/309"}],"wp:attachment":[{"href":"https:\/\/numsimlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=308"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/numsimlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=308"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/numsimlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=308"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}