节奏 · Pro/Max、选模型、成本、同类分工Cadence · Pro/Max, Model Choice, Cost, and Rivals.
选档不是看价格，是看你撞哪堵墙：额度、模型、还是并发。这一章把 Pro、Max（5× / 20×）摆开，讲清 Opus / Sonnet / Haiku 各自的活该怎么派、什么时候开思考、prompt caching 怎么帮你省钱、/fast 什么时候值。再划一条同类分工的线：Claude Code 和 Cursor / Codex / Copilot / Warp 各自强在哪，一个 builder 的组合怎么搭 —— 哪些活留在 Claude Code，哪些活切走。选对节奏，比选贵的档更省钱。Choosing a plan isn't about price — it's about which wall you hit: quota, models, or concurrency. This chapter lays out Pro and Max (5× / 20×), which work to route to Opus / Sonnet / Haiku, when to turn on thinking, how prompt caching saves money, and when /fast is worth it. Then it draws the boundary with rivals: where Claude Code, Cursor, Codex, Copilot, and Warp each win, and how a builder's stack fits together — what stays in Claude Code and what switches away. Picking the right cadence saves more than picking the priciest plan.
11 分钟 · 初稿 2026.0711 Min · Drafted 2026.07
选档的错误问法是「哪档划算」,正确问法是「我撞的是哪堵墙」。额度不够,是一堵墙;默认模型不够强,是另一堵;等不起响应速度,是第三堵 —— 三堵墙对应三个不同的开关,只有一个真需要花更多钱。这一章把档位、模型、力度、成本四个旋钮摆开,最后划一条和同类的分工线。The wrong question for picking a plan is 'which tier is the best deal'; the right one is 'which wall am I hitting.' Not enough quota is one wall; a default model that's not strong enough is another; responses too slow to wait for is a third — three walls, three different switches, and only one of them actually costs more money. This chapter lays out the four dials — plan, model, effort, cost — then draws the line of division with the rivals.
— I
档位与三堵墙Plans, and the Three Walls.
先把价目摆平:Pro 月付 $20(年付折合 $17/月,一次预付 $200);Max 从 $100 起,按比 Pro 多 5× 还是 20× 用量分两档。1注 1Note 1Claude · Pricing —— Pro:月付 $20,年付折合每月 $17(一次预付 $200);Max:从 $100 起,可选比 Pro 多 5× 或 20× 用量两档;Pro 与 Max 均标明含 Claude Code。截至 2026-07-17。Claude · Pricing — Pro: $20 billed monthly, or $17/month on annual billing ($200 up front); Max: from $100, with a choice of 5x or 20x more usage than Pro; both Pro and Max list Claude Code as included. As of 2026-07-17.两档都含 Claude Code,并且和 Claude 应用共享同一个额度池 —— 你在网页里聊掉的,和在终端里跑掉的,是同一笔钱。2注 2Note 2Claude 帮助中心 · Use Claude Code with your Pro or Max plan —— 一份订阅同时覆盖 Claude 应用与 Claude Code,两边用量共享同一个额度池。截至 2026-07-17。Claude Help Center · Use Claude Code with your Pro or Max plan — one subscription covers the Claude apps and Claude Code, with usage drawing on a shared limit pool. As of 2026-07-17.Flatten the price list first: Pro is $20 monthly ($17/month on annual billing, $200 up front); Max starts at $100, in two tiers by whether you want 5x or 20x more usage than Pro.1注 1Note 1Claude · Pricing —— Pro:月付 $20,年付折合每月 $17(一次预付 $200);Max:从 $100 起,可选比 Pro 多 5× 或 20× 用量两档;Pro 与 Max 均标明含 Claude Code。截至 2026-07-17。Claude · Pricing — Pro: $20 billed monthly, or $17/month on annual billing ($200 up front); Max: from $100, with a choice of 5x or 20x more usage than Pro; both Pro and Max list Claude Code as included. As of 2026-07-17. Both include Claude Code, and both share one usage pool with the Claude apps — what you chat away on the web and what you run through the terminal draw on the same purse.2注 2Note 2Claude 帮助中心 · Use Claude Code with your Pro or Max plan —— 一份订阅同时覆盖 Claude 应用与 Claude Code,两边用量共享同一个额度池。截至 2026-07-17。Claude Help Center · Use Claude Code with your Pro or Max plan — one subscription covers the Claude apps and Claude Code, with usage drawing on a shared limit pool. As of 2026-07-17.
你撞的墙The wall you hit
症状Symptom
开关The switch
额度Quota
下午就见底,活排着队Dry by afternoon, work queuing
升 Max 档,或先做第 III 节的省额度纪律Move up a Max tier — or apply section III's discipline first
模型Model
难题上明显吃力Visibly straining on hard problems
Pro 默认 Sonnet 5,Max 默认 Opus 4.8 且 Opus 自动 1M 上下文Pro defaults to Sonnet 5; Max to Opus 4.8, with Opus auto-1M context
速度Speed
交互时等不起Interactive waits you can't afford
/fast(走 usage credits,另一本账)/fast (usage credits — a separate ledger)
第二行值得展开:两档的差距不只是量,还有默认起点 —— Pro 账号的 default 落在 Sonnet 5,Max 落在 Opus 4.8,而且 Max / Team / Enterprise 档的 Opus 自动升到 1M token 上下文,长会话、大仓库时这一条常常比用量倍数更值钱。3注 3Note 3Claude Code Docs · Model configuration —— 默认模型按账户:Pro(及 Team Standard、Enterprise 订阅席位)默认 Sonnet 5,Max(及 Team Premium、按量 Enterprise、API)默认 Opus 4.8;Max / Team / Enterprise 档的 Opus 自动升到 1M 上下文。别名:opus / sonnet / haiku / fable(Fable 5 任何档都非默认,需 /model fable 主动选)、opusplan(计划用 Opus、执行切 Sonnet)、sonnet[1m]。effort:Sonnet 5 与 Opus 4.8 支持 low–max 五档,默认 high,/effort 调;思考由 effort 主控,Option+T(macOS)/ Alt+T 会话内开关,/config 设全局,Ctrl+O 展开思考内容。会话中途用 /model 选择器切换会弹确认(直接给定名字则立即切)—— 无论哪种,下一条回复都要无缓存重读全部历史。截至 2026-07-17。Claude Code Docs · Model configuration — default model by account: Pro (and Team Standard, Enterprise subscription seats) defaults to Sonnet 5; Max (and Team Premium, pay-as-you-go Enterprise, the API) defaults to Opus 4.8; on Max / Team / Enterprise, Opus is auto-upgraded to the 1M context window. Aliases: opus / sonnet / haiku / fable (Fable 5 is the default on no account type — select it with /model fable), opusplan (Opus to plan, Sonnet to execute), sonnet[1m]. Effort: Sonnet 5 and Opus 4.8 support five levels low–max, default high, adjusted via /effort; thinking is governed primarily by effort, toggled per session with Option+T (macOS) / Alt+T, set globally in /config, expanded with Ctrl+O. A mid-session /model switch via the picker asks for confirmation (giving a name directly switches at once) — either way the next response re-reads the full history uncached. As of 2026-07-17.判断顺序照表走:先确认撞的是哪堵墙,再动钱包 —— 三堵墙里只有额度墙必须用钱解决。The second row deserves unpacking: the tiers differ not just in volume but in starting point — a Pro account's default resolves to Sonnet 5, Max to Opus 4.8, and on Max / Team / Enterprise, Opus is auto-upgraded to the 1M-token context window, which on long sessions over large repos is often worth more than the usage multiple.3注 3Note 3Claude Code Docs · Model configuration —— 默认模型按账户:Pro(及 Team Standard、Enterprise 订阅席位)默认 Sonnet 5,Max(及 Team Premium、按量 Enterprise、API)默认 Opus 4.8;Max / Team / Enterprise 档的 Opus 自动升到 1M 上下文。别名:opus / sonnet / haiku / fable(Fable 5 任何档都非默认,需 /model fable 主动选)、opusplan(计划用 Opus、执行切 Sonnet)、sonnet[1m]。effort:Sonnet 5 与 Opus 4.8 支持 low–max 五档,默认 high,/effort 调;思考由 effort 主控,Option+T(macOS)/ Alt+T 会话内开关,/config 设全局,Ctrl+O 展开思考内容。会话中途用 /model 选择器切换会弹确认(直接给定名字则立即切)—— 无论哪种,下一条回复都要无缓存重读全部历史。截至 2026-07-17。Claude Code Docs · Model configuration — default model by account: Pro (and Team Standard, Enterprise subscription seats) defaults to Sonnet 5; Max (and Team Premium, pay-as-you-go Enterprise, the API) defaults to Opus 4.8; on Max / Team / Enterprise, Opus is auto-upgraded to the 1M context window. Aliases: opus / sonnet / haiku / fable (Fable 5 is the default on no account type — select it with /model fable), opusplan (Opus to plan, Sonnet to execute), sonnet[1m]. Effort: Sonnet 5 and Opus 4.8 support five levels low–max, default high, adjusted via /effort; thinking is governed primarily by effort, toggled per session with Option+T (macOS) / Alt+T, set globally in /config, expanded with Ctrl+O. A mid-session /model switch via the picker asks for confirmation (giving a name directly switches at once) — either way the next response re-reads the full history uncached. As of 2026-07-17. Follow the table's order: identify the wall first, then reach for the wallet — of the three, only the quota wall must be solved with money.
— II
模型与力度:两个旋钮,一套分工Model and Effort: Two Dials, One Division of Labor.
模型的日常分工用别名说话:sonnet(当前解析到 Sonnet 5)管日常编码,opus(Opus 4.8)管难题与长程任务,haiku 管机械杂活 —— 最顺手的用法不是主对话换来换去,而是第 5 章那招:主对话坐大模型,批量杂活派给 model: haiku 的 subagent。两个组合别名替你省判断:opusplan 计划时用 Opus 想、执行时切 Sonnet 干,正是第 2 章那个循环的性价比形态;还有一个天花板 fable(Fable 5),任何档都不是默认,最难的长程活才值得 /model fable 主动请它出山。3注 3Note 3Claude Code Docs · Model configuration —— 默认模型按账户:Pro(及 Team Standard、Enterprise 订阅席位)默认 Sonnet 5,Max(及 Team Premium、按量 Enterprise、API)默认 Opus 4.8;Max / Team / Enterprise 档的 Opus 自动升到 1M 上下文。别名:opus / sonnet / haiku / fable(Fable 5 任何档都非默认,需 /model fable 主动选)、opusplan(计划用 Opus、执行切 Sonnet)、sonnet[1m]。effort:Sonnet 5 与 Opus 4.8 支持 low–max 五档,默认 high,/effort 调;思考由 effort 主控,Option+T(macOS)/ Alt+T 会话内开关,/config 设全局,Ctrl+O 展开思考内容。会话中途用 /model 选择器切换会弹确认(直接给定名字则立即切)—— 无论哪种,下一条回复都要无缓存重读全部历史。截至 2026-07-17。Claude Code Docs · Model configuration — default model by account: Pro (and Team Standard, Enterprise subscription seats) defaults to Sonnet 5; Max (and Team Premium, pay-as-you-go Enterprise, the API) defaults to Opus 4.8; on Max / Team / Enterprise, Opus is auto-upgraded to the 1M context window. Aliases: opus / sonnet / haiku / fable (Fable 5 is the default on no account type — select it with /model fable), opusplan (Opus to plan, Sonnet to execute), sonnet[1m]. Effort: Sonnet 5 and Opus 4.8 support five levels low–max, default high, adjusted via /effort; thinking is governed primarily by effort, toggled per session with Option+T (macOS) / Alt+T, set globally in /config, expanded with Ctrl+O. A mid-session /model switch via the picker asks for confirmation (giving a name directly switches at once) — either way the next response re-reads the full history uncached. As of 2026-07-17.The daily division of models speaks in aliases: sonnet (currently resolving to Sonnet 5) for everyday coding, opus (Opus 4.8) for hard problems and long-horizon work, haiku for mechanical chores — and the smoothest arrangement isn't swapping your main conversation around, but chapter 5's move: the big model holds the main seat while bulk chores go to subagents with model: haiku. Two composite aliases save you the judgment: opusplan thinks in Opus during planning and switches to Sonnet to execute — the cost-effective shape of chapter 2's loop; and there's a ceiling, fable (Fable 5), default on no plan, worth summoning with /model fable only for the hardest long-horizon jobs.3注 3Note 3Claude Code Docs · Model configuration —— 默认模型按账户:Pro(及 Team Standard、Enterprise 订阅席位)默认 Sonnet 5,Max(及 Team Premium、按量 Enterprise、API)默认 Opus 4.8;Max / Team / Enterprise 档的 Opus 自动升到 1M 上下文。别名:opus / sonnet / haiku / fable(Fable 5 任何档都非默认,需 /model fable 主动选)、opusplan(计划用 Opus、执行切 Sonnet)、sonnet[1m]。effort:Sonnet 5 与 Opus 4.8 支持 low–max 五档,默认 high,/effort 调;思考由 effort 主控,Option+T(macOS)/ Alt+T 会话内开关,/config 设全局,Ctrl+O 展开思考内容。会话中途用 /model 选择器切换会弹确认(直接给定名字则立即切)—— 无论哪种,下一条回复都要无缓存重读全部历史。截至 2026-07-17。Claude Code Docs · Model configuration — default model by account: Pro (and Team Standard, Enterprise subscription seats) defaults to Sonnet 5; Max (and Team Premium, pay-as-you-go Enterprise, the API) defaults to Opus 4.8; on Max / Team / Enterprise, Opus is auto-upgraded to the 1M context window. Aliases: opus / sonnet / haiku / fable (Fable 5 is the default on no account type — select it with /model fable), opusplan (Opus to plan, Sonnet to execute), sonnet[1m]. Effort: Sonnet 5 and Opus 4.8 support five levels low–max, default high, adjusted via /effort; thinking is governed primarily by effort, toggled per session with Option+T (macOS) / Alt+T, set globally in /config, expanded with Ctrl+O. A mid-session /model switch via the picker asks for confirmation (giving a name directly switches at once) — either way the next response re-reads the full history uncached. As of 2026-07-17.第二个旋钮是力度。effort 五档(low 到 max)控制它每一步想多深,Sonnet 5 与 Opus 4.8 默认都在 high:难题升 /effort xhigh,机械活降 low —— 同一个模型,力度对了比换模型便宜。3注 3Note 3Claude Code Docs · Model configuration —— 默认模型按账户:Pro(及 Team Standard、Enterprise 订阅席位)默认 Sonnet 5,Max(及 Team Premium、按量 Enterprise、API)默认 Opus 4.8;Max / Team / Enterprise 档的 Opus 自动升到 1M 上下文。别名:opus / sonnet / haiku / fable(Fable 5 任何档都非默认,需 /model fable 主动选)、opusplan(计划用 Opus、执行切 Sonnet)、sonnet[1m]。effort:Sonnet 5 与 Opus 4.8 支持 low–max 五档,默认 high,/effort 调;思考由 effort 主控,Option+T(macOS)/ Alt+T 会话内开关,/config 设全局,Ctrl+O 展开思考内容。会话中途用 /model 选择器切换会弹确认(直接给定名字则立即切)—— 无论哪种,下一条回复都要无缓存重读全部历史。截至 2026-07-17。Claude Code Docs · Model configuration — default model by account: Pro (and Team Standard, Enterprise subscription seats) defaults to Sonnet 5; Max (and Team Premium, pay-as-you-go Enterprise, the API) defaults to Opus 4.8; on Max / Team / Enterprise, Opus is auto-upgraded to the 1M context window. Aliases: opus / sonnet / haiku / fable (Fable 5 is the default on no account type — select it with /model fable), opusplan (Opus to plan, Sonnet to execute), sonnet[1m]. Effort: Sonnet 5 and Opus 4.8 support five levels low–max, default high, adjusted via /effort; thinking is governed primarily by effort, toggled per session with Option+T (macOS) / Alt+T, set globally in /config, expanded with Ctrl+O. A mid-session /model switch via the picker asks for confirmation (giving a name directly switches at once) — either way the next response re-reads the full history uncached. As of 2026-07-17.「开思考」在这套体系里已经不是单独的开关:思考深度由 effort 主控,你只保留否决权 —— Option+T(macOS)/ Alt+T 会话内开关,/config 定全局,Ctrl+O 展开看它想了什么。3注 3Note 3Claude Code Docs · Model configuration —— 默认模型按账户:Pro(及 Team Standard、Enterprise 订阅席位)默认 Sonnet 5,Max(及 Team Premium、按量 Enterprise、API)默认 Opus 4.8;Max / Team / Enterprise 档的 Opus 自动升到 1M 上下文。别名:opus / sonnet / haiku / fable(Fable 5 任何档都非默认,需 /model fable 主动选)、opusplan(计划用 Opus、执行切 Sonnet)、sonnet[1m]。effort:Sonnet 5 与 Opus 4.8 支持 low–max 五档,默认 high,/effort 调;思考由 effort 主控,Option+T(macOS)/ Alt+T 会话内开关,/config 设全局,Ctrl+O 展开思考内容。会话中途用 /model 选择器切换会弹确认(直接给定名字则立即切)—— 无论哪种,下一条回复都要无缓存重读全部历史。截至 2026-07-17。Claude Code Docs · Model configuration — default model by account: Pro (and Team Standard, Enterprise subscription seats) defaults to Sonnet 5; Max (and Team Premium, pay-as-you-go Enterprise, the API) defaults to Opus 4.8; on Max / Team / Enterprise, Opus is auto-upgraded to the 1M context window. Aliases: opus / sonnet / haiku / fable (Fable 5 is the default on no account type — select it with /model fable), opusplan (Opus to plan, Sonnet to execute), sonnet[1m]. Effort: Sonnet 5 and Opus 4.8 support five levels low–max, default high, adjusted via /effort; thinking is governed primarily by effort, toggled per session with Option+T (macOS) / Alt+T, set globally in /config, expanded with Ctrl+O. A mid-session /model switch via the picker asks for confirmation (giving a name directly switches at once) — either way the next response re-reads the full history uncached. As of 2026-07-17.The second dial is effort. Five levels (low to max) control how deeply it thinks per step; Sonnet 5 and Opus 4.8 both default to high — raise to /effort xhigh for hard problems, drop to low for mechanical ones. Same model, right effort, cheaper than switching models.3注 3Note 3Claude Code Docs · Model configuration —— 默认模型按账户:Pro(及 Team Standard、Enterprise 订阅席位)默认 Sonnet 5,Max(及 Team Premium、按量 Enterprise、API)默认 Opus 4.8;Max / Team / Enterprise 档的 Opus 自动升到 1M 上下文。别名:opus / sonnet / haiku / fable(Fable 5 任何档都非默认,需 /model fable 主动选)、opusplan(计划用 Opus、执行切 Sonnet)、sonnet[1m]。effort:Sonnet 5 与 Opus 4.8 支持 low–max 五档,默认 high,/effort 调;思考由 effort 主控,Option+T(macOS)/ Alt+T 会话内开关,/config 设全局,Ctrl+O 展开思考内容。会话中途用 /model 选择器切换会弹确认(直接给定名字则立即切)—— 无论哪种,下一条回复都要无缓存重读全部历史。截至 2026-07-17。Claude Code Docs · Model configuration — default model by account: Pro (and Team Standard, Enterprise subscription seats) defaults to Sonnet 5; Max (and Team Premium, pay-as-you-go Enterprise, the API) defaults to Opus 4.8; on Max / Team / Enterprise, Opus is auto-upgraded to the 1M context window. Aliases: opus / sonnet / haiku / fable (Fable 5 is the default on no account type — select it with /model fable), opusplan (Opus to plan, Sonnet to execute), sonnet[1m]. Effort: Sonnet 5 and Opus 4.8 support five levels low–max, default high, adjusted via /effort; thinking is governed primarily by effort, toggled per session with Option+T (macOS) / Alt+T, set globally in /config, expanded with Ctrl+O. A mid-session /model switch via the picker asks for confirmation (giving a name directly switches at once) — either way the next response re-reads the full history uncached. As of 2026-07-17. And 'turning on thinking' is no longer its own switch in this scheme: depth is governed by effort, and you keep only the veto — Option+T (macOS) / Alt+T toggles per session, /config sets the global default, Ctrl+O expands what it thought.3注 3Note 3Claude Code Docs · Model configuration —— 默认模型按账户:Pro(及 Team Standard、Enterprise 订阅席位)默认 Sonnet 5,Max(及 Team Premium、按量 Enterprise、API)默认 Opus 4.8;Max / Team / Enterprise 档的 Opus 自动升到 1M 上下文。别名:opus / sonnet / haiku / fable(Fable 5 任何档都非默认,需 /model fable 主动选)、opusplan(计划用 Opus、执行切 Sonnet)、sonnet[1m]。effort:Sonnet 5 与 Opus 4.8 支持 low–max 五档,默认 high,/effort 调;思考由 effort 主控,Option+T(macOS)/ Alt+T 会话内开关,/config 设全局,Ctrl+O 展开思考内容。会话中途用 /model 选择器切换会弹确认(直接给定名字则立即切)—— 无论哪种,下一条回复都要无缓存重读全部历史。截至 2026-07-17。Claude Code Docs · Model configuration — default model by account: Pro (and Team Standard, Enterprise subscription seats) defaults to Sonnet 5; Max (and Team Premium, pay-as-you-go Enterprise, the API) defaults to Opus 4.8; on Max / Team / Enterprise, Opus is auto-upgraded to the 1M context window. Aliases: opus / sonnet / haiku / fable (Fable 5 is the default on no account type — select it with /model fable), opusplan (Opus to plan, Sonnet to execute), sonnet[1m]. Effort: Sonnet 5 and Opus 4.8 support five levels low–max, default high, adjusted via /effort; thinking is governed primarily by effort, toggled per session with Option+T (macOS) / Alt+T, set globally in /config, expanded with Ctrl+O. A mid-session /model switch via the picker asks for confirmation (giving a name directly switches at once) — either way the next response re-reads the full history uncached. As of 2026-07-17.
第一律:别在会话中途乱切。prompt caching 全自动打理,订阅账号还自动享受 1 小时的长缓存 —— 但缓存按模型、按 effort 分开存,会话中途 /model 或 /effort 一切,下一条回复就要无缓存重读全部历史,又慢又费;用选择器切换时 Claude Code 会弹确认框拦你,正是这个原因。官方建议照抄即可:模型和力度在会话开头定好,/compact 留给任务之间的自然断点;走错路想回头,/rewind 比 /compact 便宜 —— 它落回的是已有缓存。5注 5Note 5Claude Code Docs · How Claude Code uses prompt caching —— 缓存全自动管理;订阅账号自动用 1 小时 TTL(超出套餐、转 usage credits 后自动降到 5 分钟)。会作废缓存、让下一条又慢又贵的动作:中途切模型(每个模型各有缓存)、改 effort、首次开 fast mode、/compact;/rewind 反而落回已有缓存。官方建议:模型和 effort 在会话开头定好,/compact 留给任务间的自然断点。截至 2026-07-17。Claude Code Docs · How Claude Code uses prompt caching — caching is fully automatic; subscription accounts get the 1-hour TTL automatically (which drops to 5 minutes once you exceed the plan and draw on usage credits). Actions that invalidate it and make the next turn slow and expensive: a mid-session model switch (each model has its own cache), an effort change, the first fast-mode enable, and /compact; /rewind instead lands back on an existing cache. The official tip: pick your model and effort at the top of a session, and save /compact for natural breaks between tasks. As of 2026-07-17.Law one: don't switch mid-session. Prompt caching is managed automatically, and subscription accounts get the 1-hour long cache for free — but caches are stored per model and per effort level, so a mid-session /model or /effort makes the next reply re-read the entire history uncached, slow and costly; switching via the picker is exactly why Claude Code raises a confirmation dialog. Copy the official advice as-is: fix model and effort at the top of the session, save /compact for natural breaks between tasks — and when you want to back out of a wrong path, /rewind is cheaper than /compact: it lands on a cache that already exists.5注 5Note 5Claude Code Docs · How Claude Code uses prompt caching —— 缓存全自动管理;订阅账号自动用 1 小时 TTL(超出套餐、转 usage credits 后自动降到 5 分钟)。会作废缓存、让下一条又慢又贵的动作:中途切模型(每个模型各有缓存)、改 effort、首次开 fast mode、/compact;/rewind 反而落回已有缓存。官方建议:模型和 effort 在会话开头定好,/compact 留给任务间的自然断点。截至 2026-07-17。Claude Code Docs · How Claude Code uses prompt caching — caching is fully automatic; subscription accounts get the 1-hour TTL automatically (which drops to 5 minutes once you exceed the plan and draw on usage credits). Actions that invalidate it and make the next turn slow and expensive: a mid-session model switch (each model has its own cache), an effort change, the first fast-mode enable, and /compact; /rewind instead lands back on an existing cache. The official tip: pick your model and effort at the top of a session, and save /compact for natural breaks between tasks. As of 2026-07-17.第二律:/fast 是用钱换等待,想清楚再开。fast mode(研究预览)让同一个 Opus 最快提到 2.5 倍速,代价是 $10 / $50 每百万 token 的高价 —— 订阅用户注意,它不吃你的订阅额度,只从 usage credits 扣,是单独的一本账。4注 4Note 4Claude Code Docs · Fast mode —— 研究预览:同一个 Opus、优先速度的配置,最快 2.5×,按 MTok 计价 $10 / $50(Opus 4.8);/fast 开关;订阅档(Pro/Max/Team/Enterprise)只能经 usage credits 使用,不占订阅额度、从第一个 token 起按 fast 价计;撞到 fast 限流自动回落标准速度。截至 2026-07-17。Claude Code Docs · Fast mode — research preview: the same Opus in a speed-first configuration, up to 2.5x faster, priced $10 / $50 per MTok (Opus 4.8); toggled with /fast; on subscription plans (Pro/Max/Team/Enterprise) it's available via usage credits only — outside the plan's limits, billed at the fast rate from the first token; on hitting the fast rate limit it falls back to standard speed automatically. As of 2026-07-17.所以它的正确用法窄而清晰:赶工的交互时段开(快速迭代、现场调试),长的自动化任务与 CI 一律关;撞到 fast 的限流它会自动回落标准速度,不至于断档。4注 4Note 4Claude Code Docs · Fast mode —— 研究预览:同一个 Opus、优先速度的配置,最快 2.5×,按 MTok 计价 $10 / $50(Opus 4.8);/fast 开关;订阅档(Pro/Max/Team/Enterprise)只能经 usage credits 使用,不占订阅额度、从第一个 token 起按 fast 价计;撞到 fast 限流自动回落标准速度。截至 2026-07-17。Claude Code Docs · Fast mode — research preview: the same Opus in a speed-first configuration, up to 2.5x faster, priced $10 / $50 per MTok (Opus 4.8); toggled with /fast; on subscription plans (Pro/Max/Team/Enterprise) it's available via usage credits only — outside the plan's limits, billed at the fast rate from the first token; on hitting the fast rate limit it falls back to standard speed automatically. As of 2026-07-17.第三律:上下文纪律就是额度纪律 —— 换任务 /clear、调研派 subagent(第 5 章)、检查交给 hooks(第 7 章),每一条前面章节的习惯,在账单上都有回声。Law two: /fast trades money for waiting — decide before you flip it. Fast mode (research preview) runs the same Opus up to 2.5x faster at a premium $10 / $50 per million tokens — and subscribers note: it doesn't draw on your plan's quota, only on usage credits, a separate ledger.4注 4Note 4Claude Code Docs · Fast mode —— 研究预览:同一个 Opus、优先速度的配置,最快 2.5×,按 MTok 计价 $10 / $50(Opus 4.8);/fast 开关;订阅档(Pro/Max/Team/Enterprise)只能经 usage credits 使用,不占订阅额度、从第一个 token 起按 fast 价计;撞到 fast 限流自动回落标准速度。截至 2026-07-17。Claude Code Docs · Fast mode — research preview: the same Opus in a speed-first configuration, up to 2.5x faster, priced $10 / $50 per MTok (Opus 4.8); toggled with /fast; on subscription plans (Pro/Max/Team/Enterprise) it's available via usage credits only — outside the plan's limits, billed at the fast rate from the first token; on hitting the fast rate limit it falls back to standard speed automatically. As of 2026-07-17. Its correct use is therefore narrow and clear: on for interactive crunch (rapid iteration, live debugging), off for long automated runs and CI; when it hits the fast rate limit it falls back to standard speed on its own, so nothing stalls.4注 4Note 4Claude Code Docs · Fast mode —— 研究预览:同一个 Opus、优先速度的配置,最快 2.5×,按 MTok 计价 $10 / $50(Opus 4.8);/fast 开关;订阅档(Pro/Max/Team/Enterprise)只能经 usage credits 使用,不占订阅额度、从第一个 token 起按 fast 价计;撞到 fast 限流自动回落标准速度。截至 2026-07-17。Claude Code Docs · Fast mode — research preview: the same Opus in a speed-first configuration, up to 2.5x faster, priced $10 / $50 per MTok (Opus 4.8); toggled with /fast; on subscription plans (Pro/Max/Team/Enterprise) it's available via usage credits only — outside the plan's limits, billed at the fast rate from the first token; on hitting the fast rate limit it falls back to standard speed automatically. As of 2026-07-17. Law three: context discipline is quota discipline — /clear between tasks, subagents for research (chapter 5), hooks for checks (chapter 7): every habit from the earlier chapters echoes on the bill.
— IV
同类分工:入口跟着活走The Rivals: Entrances Follow the Work.
最后划分工线 —— 不是黑稿,是各自的主场。Cursor(个人档 Pro $20 起,顶到 Ultra $200)的主场在编辑器:你要「陪着改、盯着每一步 diff」的工作节奏时,编辑器侧的 agent 工作台仍是最顺的入口。6注 6Note 6Cursor · Pricing —— 个人档逐档(月付):Hobby 免费、Pro $20、Pro+ $60、Ultra $200;Teams $40/座·月,Enterprise 定制;页面默认的年付视图显示 Pro $16、Teams $32(八折)。逐档价格取自页面自带的结构化数据(JSON-LD)与档位选择器;卖点为 Agent 请求、code review、cloud agents。截至 2026-07-17。Cursor · Pricing — individual tiers itemized (monthly): Hobby free, Pro $20, Pro+ $60, Ultra $200; Teams $40/seat·mo, Enterprise custom; the page's default annual view shows Pro $16 and Teams $32 (20% off). Per-tier prices taken from the page's own structured data (JSON-LD) and tier selector; the pitch is Agent requests, code review, and cloud agents. As of 2026-07-17.Copilot(Pro $10 起)的主场在 GitHub 本体:issue 直接派给它、收回一个 PR,甚至能把活转派给 Claude、Codex 这些第三方 agent —— 你的团队整个活在 GitHub 流程里时,它是那条流程自带的手。7注 7Note 7GitHub Copilot · Plans —— Free / Pro $10 / Pro+ $39 / Max $100(每人·月),另有 Business 与 Enterprise;卖点原话:「把活派给 Copilot,它创建一个 pull request」,并可把任务转派给 Claude、Codex 等第三方编码 agent。截至 2026-07-17。GitHub Copilot · Plans — Free / Pro $10 / Pro+ $39 / Max $100 (per user·mo), plus Business and Enterprise; the pitch verbatim: 'Assign work to Copilot and it creates a pull request,' with the option to delegate tasks to third-party coding agents like Claude and Codex. As of 2026-07-17.Codex 是 OpenAI 的对位产品,CLI、IDE 扩展、云端、ChatGPT 应用四路入口 —— 已经付着 ChatGPT 的人自然先试它。8注 8Note 8OpenAI · Codex 文档 —— OpenAI 的编码 agent,入口覆盖 ChatGPT 桌面端与网页、Codex CLI、IDE 扩展与 Codex cloud。截至 2026-07-17。OpenAI · Codex docs — OpenAI's coding agent, with entrances across the ChatGPT desktop app and web, the Codex CLI, an IDE extension, and Codex cloud. As of 2026-07-17.Warp 自我定位是「面向 agentic coding 的现代终端」(Build $20/月):当你的活以终端会话本身为中心 —— 多面板、长命令、运维流 —— 它贴得更近。9注 9Note 9Warp · Pricing —— 自我定位「面向 agentic coding 的现代终端」;Free / Build $20 / Max $200(月付),另有 Business 与 Enterprise。截至 2026-07-17。Warp · Pricing — self-positioned as 'a modern terminal for agentic coding'; Free / Build $20 / Max $200 monthly, plus Business and Enterprise. As of 2026-07-17.组合的原则一句话:入口跟着活走,订阅别叠着买 —— 每多一份订阅都该对应一类你真在做的活;先把一份用满,再谈第二份。Now the division — no hit piece, just home turf. Cursor (individual tiers from Pro at $20 up to Ultra at $200) owns the editor: when your rhythm is 'work alongside, watch every diff,' the editor-side agent workbench remains the smoothest entrance.6注 6Note 6Cursor · Pricing —— 个人档逐档(月付):Hobby 免费、Pro $20、Pro+ $60、Ultra $200;Teams $40/座·月,Enterprise 定制;页面默认的年付视图显示 Pro $16、Teams $32(八折)。逐档价格取自页面自带的结构化数据(JSON-LD)与档位选择器;卖点为 Agent 请求、code review、cloud agents。截至 2026-07-17。Cursor · Pricing — individual tiers itemized (monthly): Hobby free, Pro $20, Pro+ $60, Ultra $200; Teams $40/seat·mo, Enterprise custom; the page's default annual view shows Pro $16 and Teams $32 (20% off). Per-tier prices taken from the page's own structured data (JSON-LD) and tier selector; the pitch is Agent requests, code review, and cloud agents. As of 2026-07-17.Copilot (Pro from $10) owns GitHub itself: assign an issue, collect a PR — it can even delegate the task onward to third-party agents like Claude and Codex — so when your team lives entirely in GitHub flow, it's the hand that flow comes with.7注 7Note 7GitHub Copilot · Plans —— Free / Pro $10 / Pro+ $39 / Max $100(每人·月),另有 Business 与 Enterprise;卖点原话:「把活派给 Copilot,它创建一个 pull request」,并可把任务转派给 Claude、Codex 等第三方编码 agent。截至 2026-07-17。GitHub Copilot · Plans — Free / Pro $10 / Pro+ $39 / Max $100 (per user·mo), plus Business and Enterprise; the pitch verbatim: 'Assign work to Copilot and it creates a pull request,' with the option to delegate tasks to third-party coding agents like Claude and Codex. As of 2026-07-17.Codex is OpenAI's counterpart, with four entrances — CLI, IDE extension, cloud, and the ChatGPT app — the natural first try for anyone already paying for ChatGPT.8注 8Note 8OpenAI · Codex 文档 —— OpenAI 的编码 agent,入口覆盖 ChatGPT 桌面端与网页、Codex CLI、IDE 扩展与 Codex cloud。截至 2026-07-17。OpenAI · Codex docs — OpenAI's coding agent, with entrances across the ChatGPT desktop app and web, the Codex CLI, an IDE extension, and Codex cloud. As of 2026-07-17.Warp positions itself as 'a modern terminal for agentic coding' (Build $20/mo): when your work centers on the terminal session itself — panes, long commands, ops flow — it sits closer.9注 9Note 9Warp · Pricing —— 自我定位「面向 agentic coding 的现代终端」;Free / Build $20 / Max $200(月付),另有 Business 与 Enterprise。截至 2026-07-17。Warp · Pricing — self-positioned as 'a modern terminal for agentic coding'; Free / Build $20 / Max $200 monthly, plus Business and Enterprise. As of 2026-07-17. The combination principle in one line: entrances follow the work; don't stack subscriptions — each one should map to a kind of work you actually do, and the first should be used to the full before a second is discussed.动手 · 用一周数据代替直觉:Hands-on · replace instinct with a week of data:
记录一周你每次「用不了」的原因:额度见底、模型吃力、还是嫌慢 —— 月底对着三堵墙的表决定动不动钱包。Log every can't-use moment for a week: quota dry, model straining, or too slow — then decide against the three-wall table whether the wallet moves.
把 subagent 的杂活降到 haiku、主对话保持默认,跑一周看额度曲线的变化。Downshift subagent chores to haiku while the main seat stays default; run a week and watch the quota curve.
只在赶工的那一小时开 /fast,完事就关 —— 月底看 usage credits 那本账,判断这个「快」值多少钱。Flip /fast on only for the crunch hour and off after — then read the usage-credits ledger at month's end to price what 'fast' was worth.
选对节奏,
比选贵的档更省钱.
The right cadence saves more
than the pricier tier.
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