挑模型 · 三个模型与思考开关Pick a Model · Three Models and the Thinking Switch.
输入框那个模型名，你大概从没换过。挑错了，要么慢且贵，要么不够聪明。这章给 Opus、Sonnet、Haiku 和「思考」开关划一条线：什么时候花 Opus，什么时候 Sonnet 就够。That model name in the input box — you probably never change it. Pick wrong and you're either slow and overpaying or not getting enough thinking. This chapter draws the line between Opus, Sonnet, Haiku, and the thinking control.
输入框右上角那个模型名，是你每次对话都在用、却几乎从没动过的开关。默认那一档不一定是最划算的那一档。That model name in the corner of the input box is the switch you use every conversation and almost never touch. The default isn't always the most economical setting.
— I
三个模型，三种代价Three Models, Three Costs.
Sonnet 是默认主力，Opus 贵在难题，Haiku 快在量大。Sonnet is the default workhorse, Opus is worth its price on hard problems, Haiku is fast at volume.三个主力模型，价格差到 5 倍：截至 2026-07，Opus 4.8 是 $5/$25、Sonnet 5 是 $3/$15、Haiku 4.5 是 $1/$5（每百万输入/输出 token）。1注 1Note 1Claude API Docs · Models overview —— 当前主力模型 Opus 4.8（$5/$25 每百万输入/输出 token，1M 上下文）、Sonnet 5（标准 $3/$15，至 2026-08-31 入门优惠 $2/$10，1M）、Haiku 4.5（$1/$5，200K）。Sonnet 5 于 2026-06-30 发布，替代 Sonnet 4.6 成为 Free / Pro 默认。截至 2026-07。Claude API Docs · Models overview — current main models Opus 4.8 ($5/$25 per Mtok in/out, 1M context), Sonnet 5 (standard $3/$15, introductory $2/$10 through 2026-08-31, 1M), Haiku 4.5 ($1/$5, 200K). Sonnet 5 launched 2026-06-30, replacing Sonnet 4.6 as the Free / Pro default. As of 2026-07.Sonnet 5 在 2026-06-30 发布，替代了 Sonnet 4.6 成为 Free 和 Pro 的默认模型。对只用网页的人，价格不体现为账单，而是额度的消耗速度 —— 同一件活，Opus 比 Sonnet 更快撞到周限额。Three main models, a 5x price spread: as of 2026-07, Opus 4.8 is $5/$25, Sonnet 5 is $3/$15, Haiku 4.5 is $1/$5 (per million input/output tokens).1注 1Note 1Claude API Docs · Models overview —— 当前主力模型 Opus 4.8（$5/$25 每百万输入/输出 token，1M 上下文）、Sonnet 5（标准 $3/$15，至 2026-08-31 入门优惠 $2/$10，1M）、Haiku 4.5（$1/$5，200K）。Sonnet 5 于 2026-06-30 发布，替代 Sonnet 4.6 成为 Free / Pro 默认。截至 2026-07。Claude API Docs · Models overview — current main models Opus 4.8 ($5/$25 per Mtok in/out, 1M context), Sonnet 5 (standard $3/$15, introductory $2/$10 through 2026-08-31, 1M), Haiku 4.5 ($1/$5, 200K). Sonnet 5 launched 2026-06-30, replacing Sonnet 4.6 as the Free / Pro default. As of 2026-07. Sonnet 5 launched on 2026-06-30, replacing Sonnet 4.6 as the default for Free and Pro. For web-only users, price shows up not as a bill but as how fast your quota drains — the same task on Opus hits the weekly limit sooner than on Sonnet.
模型Model
每百万 token（入/出）Per Mtok (in/out)
上下文Context
什么时候选它When to pick it
Opus 4.8
$5 / $25
1M
最难的推理、长代码、要一次想透的题Hardest reasoning, long code, problems to think through in one pass
Sonnet 5
$3 / $15
1M
日常主力：大多数活它就够，性价比最好（至 8 月底入门优惠 $2/$10）The daily workhorse: enough for most work, best value (intro $2/$10 through Aug)
Haiku 4.5
$1 / $5
200K
量大、要快、不烧脑的轻活High-volume, fast, low-difficulty work
默认那档（Sonnet 5）能扛掉你绝大多数活；把 Opus 留给真正难的题 —— 多步推理、长文档综合、难调的 bug，把 Haiku 留给量大求快的轻活。换模型本身不花钱，省的是额度。贵的模型买的不是上限，是下限 —— 你多花的那点额度，换的是它在难题上更少翻车。The default (Sonnet 5) carries most of your work; save Opus for the genuinely hard ones — multi-step reasoning, long-document synthesis, gnarly debugging — and Haiku for high-volume speed. Switching models costs nothing in itself; what it saves is quota. And the pricier model buys not a higher ceiling but a higher floor — the extra quota buys fewer faceplants on the hard ones.
— II
思考开关值多少What the Thinking Switch Is Worth.
模型越来越会自己决定想多久，你不用再替它按开关。Models increasingly decide how long to think on their own — you no longer need to flip a switch.Sonnet 5 和 Opus 4.8 都使用「自适应思考」（adaptive thinking）—— 模型根据任务难度自动决定要不要多想一会儿、想多久，不需要你手动开关。2注 2Note 2Claude API Docs · What's new in Sonnet 5 —— 自适应思考（adaptive thinking）默认开启，手动延长思考（extended thinking）已移除并返回 400 错误；采样参数（temperature 等）不再接受非默认值。使用新 tokenizer，同一段文字产生约 30% 更多 token。截至 2026-07。Claude API Docs · What's new in Sonnet 5 — adaptive thinking on by default, manual extended thinking removed (returns 400); sampling parameters (temperature, etc.) no longer accepted at non-default values. Uses a new tokenizer that produces ~30% more tokens for the same text. As of 2026-07.旧版 Sonnet 4.6 上的那个手动「延长思考」（Extended thinking）开关，在 Sonnet 5 上已经取消。如果你一直没动过那个开关，什么都不用改 —— 换到 Sonnet 5 之后它自己就会在该想的时候想。Sonnet 5 and Opus 4.8 both use 'adaptive thinking' — the model automatically decides whether and how long to reason harder based on the task, no manual toggle needed.2注 2Note 2Claude API Docs · What's new in Sonnet 5 —— 自适应思考（adaptive thinking）默认开启，手动延长思考（extended thinking）已移除并返回 400 错误；采样参数（temperature 等）不再接受非默认值。使用新 tokenizer，同一段文字产生约 30% 更多 token。截至 2026-07。Claude API Docs · What's new in Sonnet 5 — adaptive thinking on by default, manual extended thinking removed (returns 400); sampling parameters (temperature, etc.) no longer accepted at non-default values. Uses a new tokenizer that produces ~30% more tokens for the same text. As of 2026-07. The manual 'Extended thinking' switch from Sonnet 4.6 has been removed on Sonnet 5. If you never touched that switch, nothing changes — after moving to Sonnet 5 it simply thinks when the task calls for it.代价和以前一样：想得越多，推理越稳，额度消耗也越快。所以同一段提示，复杂题比简单题更耗额度 —— 因为模型在后台多想了。The cost equation is unchanged: more thinking means steadier reasoning and faster quota consumption. So the same prompt uses more quota on a hard problem than an easy one — because the model thought harder behind the scenes.实际操作上你需要知道的就这些：日常大多数活 Sonnet 5 默认表现就够好；真正烧脑的推理题（多步逻辑、长代码调试、数学证明）Opus 4.8 更稳 —— 它的自适应思考起点更高，不是因为你开了什么开关，是它天生就想得更深。In practice, here's what you need to know: Sonnet 5's defaults handle most everyday work well; for genuinely hard reasoning (multi-step logic, long code debugging, math proofs) Opus 4.8 is steadier — its adaptive thinking starts at a higher floor, not because you toggled anything, but because it thinks deeper by nature.
提示词Prompt让 Claude 替你定档Let Claude pick the tier
先别急着答。看一眼下面这个任务，告诉我：
1. 它该用 Haiku、Sonnet、Opus 还是 Fable？一句话说为什么（按难度 / 步数 / 时长 / 风险）。
2. 如果是 Fable，这件活是不是真的需要跑很长时间的 agent 模式？
说完你的判断，再按它直接做；如果建议的档和我现在用的不一样，提醒我去切。
任务：<把你的任务贴这里>。Don't answer yet. Look at the task below and tell me:
1. Should it run on Haiku, Sonnet, Opus, or Fable? One line on why (by difficulty / steps / duration / risk).
2. If Fable, does this task genuinely need hours-long agent mode?
State your call, then do the task by it; if the tier you'd pick differs from what I'm on, tell me to switch.
Task: <paste your task here>.
先别急着答。看一眼下面这个任务，告诉我：
1. 它该用 Haiku、Sonnet、Opus 还是 Fable？一句话说为什么（按难度 / 步数 / 时长 / 风险）。
2. 如果是 Fable，这件活是不是真的需要跑很长时间的 agent 模式？
说完你的判断，再按它直接做；如果建议的档和我现在用的不一样，提醒我去切。
任务：把这段 300 字的产品公告翻成英文，保留所有数字和产品名，语气正式。Don't answer yet. Look at the task below and tell me:
1. Should it run on Haiku, Sonnet, Opus, or Fable? One line on why (by difficulty / steps / duration / risk).
2. If Fable, does this task genuinely need hours-long agent mode?
State your call, then do the task by it; if the tier you'd pick differs from what I'm on, tell me to switch.
Task: translate this 300-word product announcement into English, keep every number and product name, formal tone.
— III
1M 上下文不是越满越好A 1M Window Isn't Better Full.
窗口大不等于该塞满。A big window doesn't mean you should fill it.Opus 4.8 与 Sonnet 5 都带 1M token 上下文，全窗口标准计价、没有倍率（Haiku 4.5 是 200K）。3注 3Note 3Claude API Docs · Pricing —— Opus 4.8 与 Sonnet 5 含完整 1M 上下文、全窗口标准计价、无倍率；Sonnet 5 使用新 tokenizer，同一段文字产生约多 30% 的 token，等价请求成本可能高于 Sonnet 4.6。Opus 4.8 另有 Fast 模式（研究预览，$10/$50）。截至 2026-07。Claude API Docs · Pricing — Opus 4.8 and Sonnet 5 include the full 1M context window at standard pricing, no multiplier; Sonnet 5 uses a new tokenizer that produces ~30% more tokens for the same text, so an equivalent request may cost more than on Sonnet 4.6. Opus 4.8 also has a Fast mode (research preview, $10/$50). As of 2026-07.有一点要注意：Sonnet 5 换了新 tokenizer，同一段文字会产生约 30% 更多的 token —— 窗口大小没变，但能装下的文字少了。日常问答感觉不到这个差别，大量上下文场景（整本书、整个代码库）可能比以前提前碰到窗口边界。Opus 4.8 and Sonnet 5 both carry a 1M-token window, billed at standard rates across the whole thing with no multiplier (Haiku 4.5 is 200K).3注 3Note 3Claude API Docs · Pricing —— Opus 4.8 与 Sonnet 5 含完整 1M 上下文、全窗口标准计价、无倍率；Sonnet 5 使用新 tokenizer，同一段文字产生约多 30% 的 token，等价请求成本可能高于 Sonnet 4.6。Opus 4.8 另有 Fast 模式（研究预览，$10/$50）。截至 2026-07。Claude API Docs · Pricing — Opus 4.8 and Sonnet 5 include the full 1M context window at standard pricing, no multiplier; Sonnet 5 uses a new tokenizer that produces ~30% more tokens for the same text, so an equivalent request may cost more than on Sonnet 4.6. Opus 4.8 also has a Fast mode (research preview, $10/$50). As of 2026-07. One wrinkle: Sonnet 5 uses a new tokenizer that produces ~30% more tokens for the same text — the window is the same size, but it holds less text. You won't notice in everyday Q&A, but heavy-context work (a whole book, a full codebase) may hit the window boundary sooner than before.什么时候真需要长上下文：一次读完整本书、整个大代码库。什么时候别：把零散资料一股脑全贴进对话 —— 那种该分段，或者交给 Project 的知识库去检索（第 4 章），让它只取相关的几页，而不是每次都重读全部。When you genuinely need long context: a whole book, a large codebase read at once. When not: dumping scattered material into the chat — that should be chunked, or handed to a Project's knowledge base for retrieval (Chapter 4) so it pulls only the relevant pages instead of re-reading everything each time.（题外一句：Opus 4.8 另有 Fast 模式（$10/$50），研究预览，主要在 API 和 Claude Code 那边 —— 网页日常用不到。）3注 3Note 3Claude API Docs · Pricing —— Opus 4.8 与 Sonnet 5 含完整 1M 上下文、全窗口标准计价、无倍率；Sonnet 5 使用新 tokenizer，同一段文字产生约多 30% 的 token，等价请求成本可能高于 Sonnet 4.6。Opus 4.8 另有 Fast 模式（研究预览，$10/$50）。截至 2026-07。Claude API Docs · Pricing — Opus 4.8 and Sonnet 5 include the full 1M context window at standard pricing, no multiplier; Sonnet 5 uses a new tokenizer that produces ~30% more tokens for the same text, so an equivalent request may cost more than on Sonnet 4.6. Opus 4.8 also has a Fast mode (research preview, $10/$50). As of 2026-07.(An aside: Opus 4.8 has a Fast mode ($10/$50), still a research preview, mostly on the API and Claude Code side — you won't encounter it in everyday web use.)3注 3Note 3Claude API Docs · Pricing —— Opus 4.8 与 Sonnet 5 含完整 1M 上下文、全窗口标准计价、无倍率；Sonnet 5 使用新 tokenizer，同一段文字产生约多 30% 的 token，等价请求成本可能高于 Sonnet 4.6。Opus 4.8 另有 Fast 模式（研究预览，$10/$50）。截至 2026-07。Claude API Docs · Pricing — Opus 4.8 and Sonnet 5 include the full 1M context window at standard pricing, no multiplier; Sonnet 5 uses a new tokenizer that produces ~30% more tokens for the same text, so an equivalent request may cost more than on Sonnet 4.6. Opus 4.8 also has a Fast mode (research preview, $10/$50). As of 2026-07.把这章的判断用到你自己最常问的问题上：Apply this chapter's judgment to the questions you ask most:
01
挑 3 个常问的问题Pick 3 questions you ask often
从你近一周的对话里挑 3 个有代表性的问题 —— 一个简单、一个中等、一个真正烧脑的。From your past week, pick 3 representative questions — one easy, one medium, one genuinely hard.
02
同一题，两个模型各跑一遍Run each on two models
每个问题分别用 Sonnet 和 Opus 跑一遍，记下答案差异、耗时，以及你愿不愿意为那点差异多花额度。Run each on both Sonnet and Opus, noting the difference in answer, the time, and whether the gap is worth the extra quota.
03
圈出「Sonnet 就够」的那几行Circle the 'Sonnet was enough' rows
大概率只有最烧脑那一题值得 Opus。这张三行表就是你之后默认该选哪个模型的依据。Most likely only the hardest one justified Opus. That three-row table is your basis for which model to default to from now on.
默认选项替你选了模型，
没替你省钱.
The default picks your model,
not your bill.
Aklman Library
— 讨论Discussion
讨论Discussion.
评论区初始化中…Initializing comments…
01 / 01
没有匹配结果No matches.
换个关键词，或按 Esc 回到页面Try another keyword, or press Esc to return