节奏 · 档位、模型、成本与同类分工Cadence · Plans, Models, Cost, and Rivals.
截至 2026-08-05,Codex 已覆盖 Free、Go 到组织档,Work 与 Codex 在可用时共用 agentic usage;计划内用量先消耗,eligible 账号再用 credits 延长。费率已按每百万 input / cached input / output token 计,Terra 与 Luna 又在 7 月审校后降价。本章把档位、Sol / Terra / Luna、多 Agent 总 token、fast mode 与 API key 两套账摆开,所有比例都写清分母。As of August 5, 2026, Codex spans Free and Go through organization plans, while Work and Codex share agentic usage where available; included usage is consumed first, then eligible accounts may extend with credits. Rates are metered per million input / cached-input / output tokens, and Terra and Luna dropped again after the July audit. This chapter separates plans, Sol / Terra / Luna, total tokens across agents, fast mode, and the API-key ledger, with an explicit denominator for every ratio.
选档的错误问法是「哪档划算」,正确问法是「我撞的是哪堵墙」。计划内用量不够,是一堵;模型不合适,是另一堵;等不起响应速度,是第三堵。三堵墙对应的开关不同,多 Agent 还会把账叠加:换模型可能省 credits,fast mode 会多吃 credits,升级或加购才抬高总量。这一章把计划、模型、力度与现行 token-based credits 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 included usage is one wall; the wrong model is another; response latency is a third. Their switches differ, and multiple agents add their ledgers together: changing models can save credits, fast mode consumes more credits, while upgrading or buying credits raises total capacity. This chapter lays out plans, models, effort, and the current token-based credit system, then draws a division of labor with the rivals.
— I
档位与三堵墙Plans, and the Three Walls.
先把商业口径摆平:定价页把 Work 与 Codex 都列入 Free、Go、Plus、Pro、Business、Enterprise、Edu,不是另买一份「Work 订阅」;Free 是 quick、Go 是 light、Plus 起更完整。价格仍是 Free $0、Go $8、Plus $20、Pro 从 $100 起(选比 Plus 多 5× 或 20× 用量)、Business $20/人·月(年付;月付 $25)。1注 1Note 1OpenAI Codex 文档 · Pricing —— 页面把 Work 与 Codex 列为 ChatGPT Free、Go、Plus、Pro、Business、Enterprise、Edu 的内含能力,共用计划用量;Free 是 quick、Go 是 light、Plus 起更完整。价格:Free $0、Go $8、Plus $20、Pro 从 $100(选比 Plus 多 5× 或 20× 用量)、Business $20/人·月(年付,月付 $25)。实际 Work 入口仍受平台、账号与 rollout 限制;API key 旁路按 API token 计费、不含 Codex cloud。截至 2026-08-05。OpenAI Codex Docs · Pricing — the page lists Work and Codex as included capabilities across ChatGPT Free, Go, Plus, Pro, Business, Enterprise, and Edu, sharing plan usage; Free is quick, Go light, and Plus upward more complete. Prices: Free $0, Go $8, Plus $20, Pro from $100 (choose 5x or 20x more usage than Plus), Business $20/user·mo annual ($25 monthly). Actual Work access still depends on platform, account, and rollout; the API-key side road uses API token billing and excludes Codex cloud. As of 2026-08-05.但内含关系不等于今天每个账号都看得到入口:Help Center 仍把 Work 的网页 / 移动端可用性写成 eligible 付费账号,桌面也要套餐已经包含;Codex 则是桌面独立体验。读者若看不到 Work,先按账号 / 平台 / rollout 待核,不要据此推断计费另有一档。10注 10Note 10OpenAI Help Center · ChatGPT Work and Codex —— Codex 是新版 ChatGPT 桌面应用里的独立体验;Work 在 eligible 付费账号的网页 / 移动端提供,桌面端仅在套餐已包含时出现。Codex 有独立入口与历史,不是普通 ChatGPT 网页 / 移动端的可选体验。入口可见性受账号与 rollout 影响。截至 2026-08-05。OpenAI Help Center · ChatGPT Work and Codex — Codex is a separate experience in the new ChatGPT desktop app; Work is available on web / mobile for eligible paid accounts and appears on desktop only when included in the plan. Codex has a separate entrance and history and is not a selectable experience in ordinary ChatGPT web / mobile. Visibility depends on account and rollout. As of 2026-08-05.API key 是另一条账:按 API token 价付费,且没有 Codex cloud。计划内含用量先吃;eligible Plus / Pro 到限后才可加购 credits。Work 与 Codex 在可用处共用一池,切个界面不会凭空多一份额度。9注 9Note 9OpenAI Help Center · Credits in ChatGPT —— 先用计划内含额度;eligible Plus / Pro 到限后可加购 credits,不用立刻升级。Work 与 Codex 在可用处共用计划用量和 credits;购买后 12 个月过期、不滚存、不可退款。实际可购性和适用功能因计划 / 地区 / rollout 而异。截至 2026-08-05。OpenAI Help Center · Credits in ChatGPT — included plan usage is consumed first; eligible Plus / Pro users can buy credits after reaching the limit instead of immediately upgrading. Where available, Work and Codex share plan usage and credits; purchases expire after 12 months, do not roll over, and are non-refundable. Purchase eligibility and supported features vary by plan / region / rollout. As of 2026-08-05.Flatten the commercial model first: the pricing page lists Work and Codex across Free, Go, Plus, Pro, Business, Enterprise, and Edu — there is no separate 'Work subscription'; Free is quick, Go light, and Plus upward more complete. Prices remain Free $0, Go $8, Plus $20, Pro from $100 (choose 5x or 20x more usage than Plus), Business $20/user·mo annual ($25 monthly).1注 1Note 1OpenAI Codex 文档 · Pricing —— 页面把 Work 与 Codex 列为 ChatGPT Free、Go、Plus、Pro、Business、Enterprise、Edu 的内含能力,共用计划用量;Free 是 quick、Go 是 light、Plus 起更完整。价格:Free $0、Go $8、Plus $20、Pro 从 $100(选比 Plus 多 5× 或 20× 用量)、Business $20/人·月(年付,月付 $25)。实际 Work 入口仍受平台、账号与 rollout 限制;API key 旁路按 API token 计费、不含 Codex cloud。截至 2026-08-05。OpenAI Codex Docs · Pricing — the page lists Work and Codex as included capabilities across ChatGPT Free, Go, Plus, Pro, Business, Enterprise, and Edu, sharing plan usage; Free is quick, Go light, and Plus upward more complete. Prices: Free $0, Go $8, Plus $20, Pro from $100 (choose 5x or 20x more usage than Plus), Business $20/user·mo annual ($25 monthly). Actual Work access still depends on platform, account, and rollout; the API-key side road uses API token billing and excludes Codex cloud. As of 2026-08-05. But being included does not mean every account can see the entrance today: Help Center still describes Work on web / mobile for eligible paid accounts and desktop only when the plan includes it; Codex is a separate desktop experience. If Work is absent, mark account / platform / rollout as pending verification rather than inferring a separate billing tier.10注 10Note 10OpenAI Help Center · ChatGPT Work and Codex —— Codex 是新版 ChatGPT 桌面应用里的独立体验;Work 在 eligible 付费账号的网页 / 移动端提供,桌面端仅在套餐已包含时出现。Codex 有独立入口与历史,不是普通 ChatGPT 网页 / 移动端的可选体验。入口可见性受账号与 rollout 影响。截至 2026-08-05。OpenAI Help Center · ChatGPT Work and Codex — Codex is a separate experience in the new ChatGPT desktop app; Work is available on web / mobile for eligible paid accounts and appears on desktop only when included in the plan. Codex has a separate entrance and history and is not a selectable experience in ordinary ChatGPT web / mobile. Visibility depends on account and rollout. As of 2026-08-05. An API key is a separate ledger: API token pricing, with no Codex cloud. Included usage comes first; eligible Plus / Pro users can buy credits after the limit. Where available, Work and Codex share a pool, so changing interfaces does not create another allowance.9注 9Note 9OpenAI Help Center · Credits in ChatGPT —— 先用计划内含额度;eligible Plus / Pro 到限后可加购 credits,不用立刻升级。Work 与 Codex 在可用处共用计划用量和 credits;购买后 12 个月过期、不滚存、不可退款。实际可购性和适用功能因计划 / 地区 / rollout 而异。截至 2026-08-05。OpenAI Help Center · Credits in ChatGPT — included plan usage is consumed first; eligible Plus / Pro users can buy credits after reaching the limit instead of immediately upgrading. Where available, Work and Codex share plan usage and credits; purchases expire after 12 months, do not roll over, and are non-refundable. Purchase eligibility and supported features vary by plan / region / rollout. As of 2026-08-05.
你撞的墙The wall you hit
症状Symptom
开关The switch
总量Capacity
计划内用量见底,新活停在门口Included usage is dry; new work waits
先降模型 / 收上下文;再考虑 credits 或升档Downshift model / context first; then credits or upgrade
模型Model
难题上明显吃力Visibly straining on hard problems
换 Sol、升思考力度(第 II 节)Switch to Sol, raise reasoning effort (section II)
速度Speed
交互时等不起Interactive waits you can't afford
/fast(走 credits,另一本账)/fast (credits — a separate ledger)
这里必须删掉一套已经过期的心算:「Plus 每 5 小时能发 Sol 15–90 条,Pro 再乘 5 或 20」。2026 年 4 月起,Plus / Pro / Business 与新 Enterprise 的 Codex 用量已迁到按实际 input / cached input / output token 折 credits;少量既有 Enterprise 账户才可能仍在 legacy per-message。8注 8Note 8OpenAI Help Center · Codex rate card —— 2026-04 起,Plus / Pro / Business 与新 Enterprise 的 Codex 用量从按消息迁到与 API token 对齐的 credits;少量既有 Enterprise 仍可能是 legacy per-message。2026-08-05 直接读取实时表确认,标准 GPT-5.6 每 100 万 token 的 credit rate:Sol 输入 125 / 缓存输入 12.5 / 输出 750;Terra 50 / 5 / 300;Luna 5 / 0.5 / 30。Ultra 没有独立单价:任务里所有 agent 的 token 都计入总量。code review 当前用 GPT-5.3-Codex 的独立费率。费率会变,以实时表为准。OpenAI Help Center · Codex rate card — beginning in April 2026, Codex usage for Plus / Pro / Business and new Enterprise moved from per-message accounting to API-token-aligned credits; a small set of existing Enterprise accounts may remain on legacy per-message billing. Directly verified against the live table on 2026-08-05, standard GPT-5.6 credit rates per 1M tokens are: Sol 125 input / 12.5 cached input / 750 output; Terra 50 / 5 / 300; Luna 5 / 0.5 / 30. Ultra has no separate unit price: tokens from every agent in a task count toward the total. Code review currently has a separate GPT-5.3-Codex rate. Rates change; use the live table.所以当前没有一张对多数人都成立的固定「每五小时消息数」表 —— 同一句 prompt,读十个文件和读一个文件不是一条消息的成本。判断顺序也该换:先看是哪种 token、哪个模型在烧,再决定收上下文、换模型、关 fast,最后才动钱包。One stale mental model has to be deleted here: 'Plus gets roughly 15–90 Sol messages per five hours, then Pro multiplies that by 5 or 20.' Beginning in April 2026, Codex usage for Plus / Pro / Business and new Enterprise moved to credits based on actual input / cached-input / output tokens; only a small set of existing Enterprise accounts may remain on legacy per-message billing.8注 8Note 8OpenAI Help Center · Codex rate card —— 2026-04 起,Plus / Pro / Business 与新 Enterprise 的 Codex 用量从按消息迁到与 API token 对齐的 credits;少量既有 Enterprise 仍可能是 legacy per-message。2026-08-05 直接读取实时表确认,标准 GPT-5.6 每 100 万 token 的 credit rate:Sol 输入 125 / 缓存输入 12.5 / 输出 750;Terra 50 / 5 / 300;Luna 5 / 0.5 / 30。Ultra 没有独立单价:任务里所有 agent 的 token 都计入总量。code review 当前用 GPT-5.3-Codex 的独立费率。费率会变,以实时表为准。OpenAI Help Center · Codex rate card — beginning in April 2026, Codex usage for Plus / Pro / Business and new Enterprise moved from per-message accounting to API-token-aligned credits; a small set of existing Enterprise accounts may remain on legacy per-message billing. Directly verified against the live table on 2026-08-05, standard GPT-5.6 credit rates per 1M tokens are: Sol 125 input / 12.5 cached input / 750 output; Terra 50 / 5 / 300; Luna 5 / 0.5 / 30. Ultra has no separate unit price: tokens from every agent in a task count toward the total. Code review currently has a separate GPT-5.3-Codex rate. Rates change; use the live table. There is therefore no fixed 'messages per five hours' table that applies to most users — one prompt reading ten files doesn't cost what one reading a single file does. The diagnostic order changes too: see which token type and model are burning, then trim context, switch models, turn off fast, and only then move the wallet.
— II
三兄弟与力度:两个旋钮The Trio and the Effort Dial: Two Knobs.
模型的日常分工是三个 GPT-5.6 兄弟。Sol 管复杂、开放、高价值的活(旗舰,拿不准就用它);Terra 是务实全能的日常主力;Luna 跑清晰可重复、量大的活:抽取、分类、转换。2注 2Note 2OpenAI Codex 文档 · Models —— Codex 有三个 GPT-5.6 模型:Sol(复杂开放 / 高价值任务旗舰)、Terra(日常主力)、Luna(清晰可重复、量大的抽取 / 分类 / 转换)。默认 Power = Sol + medium。力度:Light(应用)/ Low(CLI)、Medium、High、Extra High;Max 给单任务更多推理,Ultra 会在 eligible Work 账号里主动用 subagents 并行拆解。gpt-5.2 与 gpt-5.3-codex 对 ChatGPT 登录的一般任务已弃用,但 GitHub / cloud code review 当前仍按 GPT-5.3-Codex 计费。云端对话默认模型不可改。截至 2026-08-05。OpenAI Codex Docs · Models — Codex has three GPT-5.6 models: Sol (flagship for complex / open-ended, high-value work), Terra (everyday workhorse), and Luna (clear, repeatable, high-volume extraction / classification / transformation). Default Power = Sol + medium. Effort: Light (app) / Low (CLI), Medium, High, Extra High; Max gives one task more reasoning, while Ultra proactively uses subagents in eligible Work accounts. gpt-5.2 and gpt-5.3-codex are deprecated for general ChatGPT-sign-in work, but GitHub / cloud code review is currently still rated as GPT-5.3-Codex. The cloud-chat default model can't be changed. As of 2026-08-05.默认 Power 就是 Sol 加 medium 推理。最省事的路由是:架构、疑难 bug、最终打磨留 Sol;常规实现先 Terra;批量读 / 扫 / 整理先 Luna。云端对话的默认模型当前不能改,所以这套路由主要用于本地 app / CLI / IDE 与你配置的 subagent roles,别在 cloud UI 里找一个不存在的切换器。2注 2Note 2OpenAI Codex 文档 · Models —— Codex 有三个 GPT-5.6 模型:Sol(复杂开放 / 高价值任务旗舰)、Terra(日常主力)、Luna(清晰可重复、量大的抽取 / 分类 / 转换)。默认 Power = Sol + medium。力度:Light(应用)/ Low(CLI)、Medium、High、Extra High;Max 给单任务更多推理,Ultra 会在 eligible Work 账号里主动用 subagents 并行拆解。gpt-5.2 与 gpt-5.3-codex 对 ChatGPT 登录的一般任务已弃用,但 GitHub / cloud code review 当前仍按 GPT-5.3-Codex 计费。云端对话默认模型不可改。截至 2026-08-05。OpenAI Codex Docs · Models — Codex has three GPT-5.6 models: Sol (flagship for complex / open-ended, high-value work), Terra (everyday workhorse), and Luna (clear, repeatable, high-volume extraction / classification / transformation). Default Power = Sol + medium. Effort: Light (app) / Low (CLI), Medium, High, Extra High; Max gives one task more reasoning, while Ultra proactively uses subagents in eligible Work accounts. gpt-5.2 and gpt-5.3-codex are deprecated for general ChatGPT-sign-in work, but GitHub / cloud code review is currently still rated as GPT-5.3-Codex. The cloud-chat default model can't be changed. As of 2026-08-05.The daily division is three GPT-5.6 siblings. Sol handles complex, open-ended, high-value work (the flagship; use it when unsure); Terra is the pragmatic everyday workhorse; Luna handles clear, repeatable, high-volume extraction, classification, and transformation.2注 2Note 2OpenAI Codex 文档 · Models —— Codex 有三个 GPT-5.6 模型:Sol(复杂开放 / 高价值任务旗舰)、Terra(日常主力)、Luna(清晰可重复、量大的抽取 / 分类 / 转换)。默认 Power = Sol + medium。力度:Light(应用)/ Low(CLI)、Medium、High、Extra High;Max 给单任务更多推理,Ultra 会在 eligible Work 账号里主动用 subagents 并行拆解。gpt-5.2 与 gpt-5.3-codex 对 ChatGPT 登录的一般任务已弃用,但 GitHub / cloud code review 当前仍按 GPT-5.3-Codex 计费。云端对话默认模型不可改。截至 2026-08-05。OpenAI Codex Docs · Models — Codex has three GPT-5.6 models: Sol (flagship for complex / open-ended, high-value work), Terra (everyday workhorse), and Luna (clear, repeatable, high-volume extraction / classification / transformation). Default Power = Sol + medium. Effort: Light (app) / Low (CLI), Medium, High, Extra High; Max gives one task more reasoning, while Ultra proactively uses subagents in eligible Work accounts. gpt-5.2 and gpt-5.3-codex are deprecated for general ChatGPT-sign-in work, but GitHub / cloud code review is currently still rated as GPT-5.3-Codex. The cloud-chat default model can't be changed. As of 2026-08-05. Default Power is Sol plus medium reasoning. The simplest routing: architecture, hard bugs, and final polish stay on Sol; routine implementation starts on Terra; batched reading / scanning / tidying starts on Luna. A cloud chat's default model currently can't be changed, so this routing mainly applies to the local app / CLI / IDE and configured subagent roles — don't hunt for a switch that isn't in the cloud UI.2注 2Note 2OpenAI Codex 文档 · Models —— Codex 有三个 GPT-5.6 模型:Sol(复杂开放 / 高价值任务旗舰)、Terra(日常主力)、Luna(清晰可重复、量大的抽取 / 分类 / 转换)。默认 Power = Sol + medium。力度:Light(应用)/ Low(CLI)、Medium、High、Extra High;Max 给单任务更多推理,Ultra 会在 eligible Work 账号里主动用 subagents 并行拆解。gpt-5.2 与 gpt-5.3-codex 对 ChatGPT 登录的一般任务已弃用,但 GitHub / cloud code review 当前仍按 GPT-5.3-Codex 计费。云端对话默认模型不可改。截至 2026-08-05。OpenAI Codex Docs · Models — Codex has three GPT-5.6 models: Sol (flagship for complex / open-ended, high-value work), Terra (everyday workhorse), and Luna (clear, repeatable, high-volume extraction / classification / transformation). Default Power = Sol + medium. Effort: Light (app) / Low (CLI), Medium, High, Extra High; Max gives one task more reasoning, while Ultra proactively uses subagents in eligible Work accounts. gpt-5.2 and gpt-5.3-codex are deprecated for general ChatGPT-sign-in work, but GitHub / cloud code review is currently still rated as GPT-5.3-Codex. The cloud-chat default model can't be changed. As of 2026-08-05.第二个旋钮是推理力度:从 Light(应用)/ Low(CLI)、Medium、到 High、Extra High,控制它每一步想多深 —— 难题升档、机械活降档,同一个模型,力度对了常比换模型便宜。2注 2Note 2OpenAI Codex 文档 · Models —— Codex 有三个 GPT-5.6 模型:Sol(复杂开放 / 高价值任务旗舰)、Terra(日常主力)、Luna(清晰可重复、量大的抽取 / 分类 / 转换)。默认 Power = Sol + medium。力度:Light(应用)/ Low(CLI)、Medium、High、Extra High;Max 给单任务更多推理,Ultra 会在 eligible Work 账号里主动用 subagents 并行拆解。gpt-5.2 与 gpt-5.3-codex 对 ChatGPT 登录的一般任务已弃用,但 GitHub / cloud code review 当前仍按 GPT-5.3-Codex 计费。云端对话默认模型不可改。截至 2026-08-05。OpenAI Codex Docs · Models — Codex has three GPT-5.6 models: Sol (flagship for complex / open-ended, high-value work), Terra (everyday workhorse), and Luna (clear, repeatable, high-volume extraction / classification / transformation). Default Power = Sol + medium. Effort: Light (app) / Low (CLI), Medium, High, Extra High; Max gives one task more reasoning, while Ultra proactively uses subagents in eligible Work accounts. gpt-5.2 and gpt-5.3-codex are deprecated for general ChatGPT-sign-in work, but GitHub / cloud code review is currently still rated as GPT-5.3-Codex. The cloud-chat default model can't be changed. As of 2026-08-05.再上面有两个特殊档:Max 给单个任务更多推理时间;Ultra 会在 eligible Work 账号里主动用 subagents 并行拆解能分块的复杂活 —— 多数活两个都用不上,入口可见性也要按账号待核。2注 2Note 2OpenAI Codex 文档 · Models —— Codex 有三个 GPT-5.6 模型:Sol(复杂开放 / 高价值任务旗舰)、Terra(日常主力)、Luna(清晰可重复、量大的抽取 / 分类 / 转换)。默认 Power = Sol + medium。力度:Light(应用)/ Low(CLI)、Medium、High、Extra High;Max 给单任务更多推理,Ultra 会在 eligible Work 账号里主动用 subagents 并行拆解。gpt-5.2 与 gpt-5.3-codex 对 ChatGPT 登录的一般任务已弃用,但 GitHub / cloud code review 当前仍按 GPT-5.3-Codex 计费。云端对话默认模型不可改。截至 2026-08-05。OpenAI Codex Docs · Models — Codex has three GPT-5.6 models: Sol (flagship for complex / open-ended, high-value work), Terra (everyday workhorse), and Luna (clear, repeatable, high-volume extraction / classification / transformation). Default Power = Sol + medium. Effort: Light (app) / Low (CLI), Medium, High, Extra High; Max gives one task more reasoning, while Ultra proactively uses subagents in eligible Work accounts. gpt-5.2 and gpt-5.3-codex are deprecated for general ChatGPT-sign-in work, but GitHub / cloud code review is currently still rated as GPT-5.3-Codex. The cloud-chat default model can't be changed. As of 2026-08-05.Ultra 不是包价:它和你手动开的子线程一样,所有 agent 的 token 都进总账。需要可控路由时,用第 6 章的 default_subagent_model 和窄角色,别把每个探索线程都固定成 Sol。还有 Pro 专属的 GPT-5.3-Codex-Spark(研究预览):near-instant 的纯文本迭代,是独立模型、有自己的限额,和 fast mode 不是一回事。3注 3Note 3OpenAI Codex 文档 · Speed —— Fast mode 把支持的模型提速 1.5×,按更高倍率吃 credits:GPT-5.6 / 5.5 为标准的 2.5×,GPT-5.4 为 2×;/fast on|off|status 切,或 config 里 service_tier = "fast"。仅 ChatGPT 登录可用(是 credit 特性);API key 走 API token 价、无 credit 倍率。GPT-5.3-Codex-Spark 是另一个更快、能力较弱的独立模型(研究预览,仅 ChatGPT Pro),有自己的用量限额,和 fast mode 不是一回事。截至 2026-08-05。OpenAI Codex Docs · Speed — Fast mode speeds a supported model up by 1.5x, consuming credits at a higher multiplier: 2.5x the standard rate for GPT-5.6 / 5.5, 2x for GPT-5.4; toggle with /fast on|off|status or service_tier = "fast" in config. Available only with ChatGPT sign-in (it's a credit feature); an API key uses API token pricing with no credit multiplier. GPT-5.3-Codex-Spark is a separate faster, less-capable model (research preview, ChatGPT Pro only) with its own limits, distinct from fast mode. As of 2026-08-05.The second knob is reasoning effort: from Light (app) / Low (CLI), Medium, up to High, Extra High, controlling how deeply it thinks per step — raise it for hard problems, drop it for mechanical ones; same model, right effort, often cheaper than switching models.2注 2Note 2OpenAI Codex 文档 · Models —— Codex 有三个 GPT-5.6 模型:Sol(复杂开放 / 高价值任务旗舰)、Terra(日常主力)、Luna(清晰可重复、量大的抽取 / 分类 / 转换)。默认 Power = Sol + medium。力度:Light(应用)/ Low(CLI)、Medium、High、Extra High;Max 给单任务更多推理,Ultra 会在 eligible Work 账号里主动用 subagents 并行拆解。gpt-5.2 与 gpt-5.3-codex 对 ChatGPT 登录的一般任务已弃用,但 GitHub / cloud code review 当前仍按 GPT-5.3-Codex 计费。云端对话默认模型不可改。截至 2026-08-05。OpenAI Codex Docs · Models — Codex has three GPT-5.6 models: Sol (flagship for complex / open-ended, high-value work), Terra (everyday workhorse), and Luna (clear, repeatable, high-volume extraction / classification / transformation). Default Power = Sol + medium. Effort: Light (app) / Low (CLI), Medium, High, Extra High; Max gives one task more reasoning, while Ultra proactively uses subagents in eligible Work accounts. gpt-5.2 and gpt-5.3-codex are deprecated for general ChatGPT-sign-in work, but GitHub / cloud code review is currently still rated as GPT-5.3-Codex. The cloud-chat default model can't be changed. As of 2026-08-05. Above that are two special settings: Max gives one task more reasoning time; Ultra proactively uses subagents in eligible Work accounts to split divisible complex work — most work needs neither, and entrance visibility remains account-dependent.2注 2Note 2OpenAI Codex 文档 · Models —— Codex 有三个 GPT-5.6 模型:Sol(复杂开放 / 高价值任务旗舰)、Terra(日常主力)、Luna(清晰可重复、量大的抽取 / 分类 / 转换)。默认 Power = Sol + medium。力度:Light(应用)/ Low(CLI)、Medium、High、Extra High;Max 给单任务更多推理,Ultra 会在 eligible Work 账号里主动用 subagents 并行拆解。gpt-5.2 与 gpt-5.3-codex 对 ChatGPT 登录的一般任务已弃用,但 GitHub / cloud code review 当前仍按 GPT-5.3-Codex 计费。云端对话默认模型不可改。截至 2026-08-05。OpenAI Codex Docs · Models — Codex has three GPT-5.6 models: Sol (flagship for complex / open-ended, high-value work), Terra (everyday workhorse), and Luna (clear, repeatable, high-volume extraction / classification / transformation). Default Power = Sol + medium. Effort: Light (app) / Low (CLI), Medium, High, Extra High; Max gives one task more reasoning, while Ultra proactively uses subagents in eligible Work accounts. gpt-5.2 and gpt-5.3-codex are deprecated for general ChatGPT-sign-in work, but GitHub / cloud code review is currently still rated as GPT-5.3-Codex. The cloud-chat default model can't be changed. As of 2026-08-05.Ultra is not flat-rate: as with manually created child threads, every agent's tokens enter the total. For controlled routing, use chapter 6's default_subagent_model and narrow roles rather than pinning every exploration thread to Sol. Pro also has GPT-5.3-Codex-Spark (research preview): near-instant text-only iteration, a separate model with its own limit, not fast mode.3注 3Note 3OpenAI Codex 文档 · Speed —— Fast mode 把支持的模型提速 1.5×,按更高倍率吃 credits:GPT-5.6 / 5.5 为标准的 2.5×,GPT-5.4 为 2×;/fast on|off|status 切,或 config 里 service_tier = "fast"。仅 ChatGPT 登录可用(是 credit 特性);API key 走 API token 价、无 credit 倍率。GPT-5.3-Codex-Spark 是另一个更快、能力较弱的独立模型(研究预览,仅 ChatGPT Pro),有自己的用量限额,和 fast mode 不是一回事。截至 2026-08-05。OpenAI Codex Docs · Speed — Fast mode speeds a supported model up by 1.5x, consuming credits at a higher multiplier: 2.5x the standard rate for GPT-5.6 / 5.5, 2x for GPT-5.4; toggle with /fast on|off|status or service_tier = "fast" in config. Available only with ChatGPT sign-in (it's a credit feature); an API key uses API token pricing with no credit multiplier. GPT-5.3-Codex-Spark is a separate faster, less-capable model (research preview, ChatGPT Pro only) with its own limits, distinct from fast mode. As of 2026-08-05.
— III
成本:credits 按哪种 token 算Cost: Which Token Type Credits Count.
credits 不是「一条消息一个价」,也不是把所有 token 模糊揉成一团。现行标准 GPT-5.6 rate card 按模型 × token 类型分别算;下面每个数字的分母都是 100 万 token,单位是 credits:8注 8Note 8OpenAI Help Center · Codex rate card —— 2026-04 起,Plus / Pro / Business 与新 Enterprise 的 Codex 用量从按消息迁到与 API token 对齐的 credits;少量既有 Enterprise 仍可能是 legacy per-message。2026-08-05 直接读取实时表确认,标准 GPT-5.6 每 100 万 token 的 credit rate:Sol 输入 125 / 缓存输入 12.5 / 输出 750;Terra 50 / 5 / 300;Luna 5 / 0.5 / 30。Ultra 没有独立单价:任务里所有 agent 的 token 都计入总量。code review 当前用 GPT-5.3-Codex 的独立费率。费率会变,以实时表为准。OpenAI Help Center · Codex rate card — beginning in April 2026, Codex usage for Plus / Pro / Business and new Enterprise moved from per-message accounting to API-token-aligned credits; a small set of existing Enterprise accounts may remain on legacy per-message billing. Directly verified against the live table on 2026-08-05, standard GPT-5.6 credit rates per 1M tokens are: Sol 125 input / 12.5 cached input / 750 output; Terra 50 / 5 / 300; Luna 5 / 0.5 / 30. Ultra has no separate unit price: tokens from every agent in a task count toward the total. Code review currently has a separate GPT-5.3-Codex rate. Rates change; use the live table.Credits aren't 'one price per message,' nor do they blur every token into one bucket. The current standard GPT-5.6 rate card prices model × token type separately; every number below is credits per 1M tokens:8注 8Note 8OpenAI Help Center · Codex rate card —— 2026-04 起,Plus / Pro / Business 与新 Enterprise 的 Codex 用量从按消息迁到与 API token 对齐的 credits;少量既有 Enterprise 仍可能是 legacy per-message。2026-08-05 直接读取实时表确认,标准 GPT-5.6 每 100 万 token 的 credit rate:Sol 输入 125 / 缓存输入 12.5 / 输出 750;Terra 50 / 5 / 300;Luna 5 / 0.5 / 30。Ultra 没有独立单价:任务里所有 agent 的 token 都计入总量。code review 当前用 GPT-5.3-Codex 的独立费率。费率会变,以实时表为准。OpenAI Help Center · Codex rate card — beginning in April 2026, Codex usage for Plus / Pro / Business and new Enterprise moved from per-message accounting to API-token-aligned credits; a small set of existing Enterprise accounts may remain on legacy per-message billing. Directly verified against the live table on 2026-08-05, standard GPT-5.6 credit rates per 1M tokens are: Sol 125 input / 12.5 cached input / 750 output; Terra 50 / 5 / 300; Luna 5 / 0.5 / 30. Ultra has no separate unit price: tokens from every agent in a task count toward the total. Code review currently has a separate GPT-5.3-Codex rate. Rates change; use the live table.
模型Model
输入 / 100 万Input / 1M
缓存输入 / 100 万Cached input / 1M
输出 / 100 万Output / 1M
Sol
125
12.50
750
Terra
50
5
300
Luna
5
0.50
30
同一 token 类型下,Terra 的当前费率水平是 Sol 的 40%(便宜 60%),Luna 是 Sol 的 4%(便宜 96%);分母始终是 100 万同类 token。输出又是同模型输入的 6 倍,缓存输入是普通输入的十分之一。8注 8Note 8OpenAI Help Center · Codex rate card —— 2026-04 起,Plus / Pro / Business 与新 Enterprise 的 Codex 用量从按消息迁到与 API token 对齐的 credits;少量既有 Enterprise 仍可能是 legacy per-message。2026-08-05 直接读取实时表确认,标准 GPT-5.6 每 100 万 token 的 credit rate:Sol 输入 125 / 缓存输入 12.5 / 输出 750;Terra 50 / 5 / 300;Luna 5 / 0.5 / 30。Ultra 没有独立单价:任务里所有 agent 的 token 都计入总量。code review 当前用 GPT-5.3-Codex 的独立费率。费率会变,以实时表为准。OpenAI Help Center · Codex rate card — beginning in April 2026, Codex usage for Plus / Pro / Business and new Enterprise moved from per-message accounting to API-token-aligned credits; a small set of existing Enterprise accounts may remain on legacy per-message billing. Directly verified against the live table on 2026-08-05, standard GPT-5.6 credit rates per 1M tokens are: Sol 125 input / 12.5 cached input / 750 output; Terra 50 / 5 / 300; Luna 5 / 0.5 / 30. Ultra has no separate unit price: tokens from every agent in a task count toward the total. Code review currently has a separate GPT-5.3-Codex rate. Rates change; use the live table.别把水平差当成一段时间里的降幅:和本书 2026-07-17 的实测表相比,Terra 三类单价都从 62.5 / 6.25 / 375 降到当前 50 / 5 / 300,即各降 20%;Luna 从 25 / 2.5 / 150 降到 5 / 0.5 / 30,即各降 80%。这是以 7 月 17 日旧价为分母的变化比较。费率会变,读到时先看实时表。Within the same token type, Terra's current rate level is 40% of Sol (60% cheaper), while Luna is 4% of Sol (96% cheaper); the denominator remains 1M tokens of the same type. Output is 6x input on each model, while cached input is one-tenth of ordinary input.8注 8Note 8OpenAI Help Center · Codex rate card —— 2026-04 起,Plus / Pro / Business 与新 Enterprise 的 Codex 用量从按消息迁到与 API token 对齐的 credits;少量既有 Enterprise 仍可能是 legacy per-message。2026-08-05 直接读取实时表确认,标准 GPT-5.6 每 100 万 token 的 credit rate:Sol 输入 125 / 缓存输入 12.5 / 输出 750;Terra 50 / 5 / 300;Luna 5 / 0.5 / 30。Ultra 没有独立单价:任务里所有 agent 的 token 都计入总量。code review 当前用 GPT-5.3-Codex 的独立费率。费率会变,以实时表为准。OpenAI Help Center · Codex rate card — beginning in April 2026, Codex usage for Plus / Pro / Business and new Enterprise moved from per-message accounting to API-token-aligned credits; a small set of existing Enterprise accounts may remain on legacy per-message billing. Directly verified against the live table on 2026-08-05, standard GPT-5.6 credit rates per 1M tokens are: Sol 125 input / 12.5 cached input / 750 output; Terra 50 / 5 / 300; Luna 5 / 0.5 / 30. Ultra has no separate unit price: tokens from every agent in a task count toward the total. Code review currently has a separate GPT-5.3-Codex rate. Rates change; use the live table. Do not confuse a level gap with a change over time: compared with this book's live check on 2026-07-17, Terra's three rates moved from 62.5 / 6.25 / 375 to 50 / 5 / 300, a 20% decrease in each; Luna moved from 25 / 2.5 / 150 to 5 / 0.5 / 30, an 80% decrease in each. Those changes use the July 17 old price as denominator. Rates change; check the live table when reading.多 Agent 的账最容易被「并行更快」遮住。主线程、explorer、worker、reviewer 各自读上下文、推理、回摘要;rate card 明确说 Ultra 任务里的所有 agent token 都计入总量,手动子线程也该按同一总账心算。8注 8Note 8OpenAI Help Center · Codex rate card —— 2026-04 起,Plus / Pro / Business 与新 Enterprise 的 Codex 用量从按消息迁到与 API token 对齐的 credits;少量既有 Enterprise 仍可能是 legacy per-message。2026-08-05 直接读取实时表确认,标准 GPT-5.6 每 100 万 token 的 credit rate:Sol 输入 125 / 缓存输入 12.5 / 输出 750;Terra 50 / 5 / 300;Luna 5 / 0.5 / 30。Ultra 没有独立单价:任务里所有 agent 的 token 都计入总量。code review 当前用 GPT-5.3-Codex 的独立费率。费率会变,以实时表为准。OpenAI Help Center · Codex rate card — beginning in April 2026, Codex usage for Plus / Pro / Business and new Enterprise moved from per-message accounting to API-token-aligned credits; a small set of existing Enterprise accounts may remain on legacy per-message billing. Directly verified against the live table on 2026-08-05, standard GPT-5.6 credit rates per 1M tokens are: Sol 125 input / 12.5 cached input / 750 output; Terra 50 / 5 / 300; Luna 5 / 0.5 / 30. Ultra has no separate unit price: tokens from every agent in a task count toward the total. Code review currently has a separate GPT-5.3-Codex rate. Rates change; use the live table.并发缩短的是墙上时间,不保证降低 token。第 6 章的省钱顺序因此很具体:读密集型探索可并行,但限定目录与问题;子 agent 回结论与证据,不抄整份日志;写密集型只留一个 owner;用 max_concurrent_threads_per_session 限并发,再用 Luna / Terra 做适合的默认子 agent,把 Sol 留给架构与最终裁决。The multi-agent ledger is easy to miss behind 'parallel is faster.' The main thread, explorer, worker, and reviewer each read context, reason, and return a summary; the rate card explicitly says every agent token in an Ultra task counts toward the total, and manual child threads deserve the same mental model.8注 8Note 8OpenAI Help Center · Codex rate card —— 2026-04 起,Plus / Pro / Business 与新 Enterprise 的 Codex 用量从按消息迁到与 API token 对齐的 credits;少量既有 Enterprise 仍可能是 legacy per-message。2026-08-05 直接读取实时表确认,标准 GPT-5.6 每 100 万 token 的 credit rate:Sol 输入 125 / 缓存输入 12.5 / 输出 750;Terra 50 / 5 / 300;Luna 5 / 0.5 / 30。Ultra 没有独立单价:任务里所有 agent 的 token 都计入总量。code review 当前用 GPT-5.3-Codex 的独立费率。费率会变,以实时表为准。OpenAI Help Center · Codex rate card — beginning in April 2026, Codex usage for Plus / Pro / Business and new Enterprise moved from per-message accounting to API-token-aligned credits; a small set of existing Enterprise accounts may remain on legacy per-message billing. Directly verified against the live table on 2026-08-05, standard GPT-5.6 credit rates per 1M tokens are: Sol 125 input / 12.5 cached input / 750 output; Terra 50 / 5 / 300; Luna 5 / 0.5 / 30. Ultra has no separate unit price: tokens from every agent in a task count toward the total. Code review currently has a separate GPT-5.3-Codex rate. Rates change; use the live table. Concurrency reduces wall-clock time; it does not guarantee fewer tokens. Chapter 6 therefore gives a concrete saving order: parallelize read-heavy exploration but constrain directories and questions; return findings and evidence rather than whole logs; keep one owner for write-heavy work; cap concurrency with max_concurrent_threads_per_session; route suitable child work to Luna / Terra and reserve Sol for architecture and final judgment.支付顺序也别倒过来:先吃计划内含用量;eligible Plus / Pro 到限后才加购 credits,不必立刻升档。买到的 credits 在支持的 agentic 功能间共用,12 个月过期、不滚存;能不能买、能花在哪还受计划 / 地区 / rollout 影响。9注 9Note 9OpenAI Help Center · Credits in ChatGPT —— 先用计划内含额度;eligible Plus / Pro 到限后可加购 credits,不用立刻升级。Work 与 Codex 在可用处共用计划用量和 credits;购买后 12 个月过期、不滚存、不可退款。实际可购性和适用功能因计划 / 地区 / rollout 而异。截至 2026-08-05。OpenAI Help Center · Credits in ChatGPT — included plan usage is consumed first; eligible Plus / Pro users can buy credits after reaching the limit instead of immediately upgrading. Where available, Work and Codex share plan usage and credits; purchases expire after 12 months, do not roll over, and are non-refundable. Purchase eligibility and supported features vary by plan / region / rollout. As of 2026-08-05.少量 legacy Enterprise 若仍按消息计,上表不直接套用 —— 看你组织后台的实际计费口径。Don't reverse the payment order either: included plan usage comes first; eligible Plus / Pro users buy credits after reaching the limit rather than upgrading immediately. Purchased credits are shared across supported agentic features, expire after 12 months, and don't roll over; purchase eligibility and supported uses still depend on plan / region / rollout.9注 9Note 9OpenAI Help Center · Credits in ChatGPT —— 先用计划内含额度;eligible Plus / Pro 到限后可加购 credits,不用立刻升级。Work 与 Codex 在可用处共用计划用量和 credits;购买后 12 个月过期、不滚存、不可退款。实际可购性和适用功能因计划 / 地区 / rollout 而异。截至 2026-08-05。OpenAI Help Center · Credits in ChatGPT — included plan usage is consumed first; eligible Plus / Pro users can buy credits after reaching the limit instead of immediately upgrading. Where available, Work and Codex share plan usage and credits; purchases expire after 12 months, do not roll over, and are non-refundable. Purchase eligibility and supported features vary by plan / region / rollout. As of 2026-08-05. If a legacy Enterprise account is still message-based, this table doesn't apply directly — use the billing basis in the organization console.Fast mode 是用 credits 换等待,想清楚再开。以同一个 GPT-5.6 模型的 standard mode 为基准,官方平均速度约 1.5×(即快约 50%),credit 消耗却是 2.5×(即多 150%);/fast on 切。3注 3Note 3OpenAI Codex 文档 · Speed —— Fast mode 把支持的模型提速 1.5×,按更高倍率吃 credits:GPT-5.6 / 5.5 为标准的 2.5×,GPT-5.4 为 2×;/fast on|off|status 切,或 config 里 service_tier = "fast"。仅 ChatGPT 登录可用(是 credit 特性);API key 走 API token 价、无 credit 倍率。GPT-5.3-Codex-Spark 是另一个更快、能力较弱的独立模型(研究预览,仅 ChatGPT Pro),有自己的用量限额,和 fast mode 不是一回事。截至 2026-08-05。OpenAI Codex Docs · Speed — Fast mode speeds a supported model up by 1.5x, consuming credits at a higher multiplier: 2.5x the standard rate for GPT-5.6 / 5.5, 2x for GPT-5.4; toggle with /fast on|off|status or service_tier = "fast" in config. Available only with ChatGPT sign-in (it's a credit feature); an API key uses API token pricing with no credit multiplier. GPT-5.3-Codex-Spark is a separate faster, less-capable model (research preview, ChatGPT Pro only) with its own limits, distinct from fast mode. As of 2026-08-05.这是 ChatGPT 登录下的 credit 特性;API key 走 API token 价,没有这笔 credit 倍率。正确用法窄而清晰:赶工的一小时开,长自动化与 CI 关 —— 速度墙也会花钱,只是花法不同于总量墙。Fast mode trades credits for waiting, so choose deliberately. Against the same GPT-5.6 model in standard mode, the official average is about 1.5x speed (roughly 50% faster) while credit consumption is 2.5x (150% higher); toggle with /fast on.3注 3Note 3OpenAI Codex 文档 · Speed —— Fast mode 把支持的模型提速 1.5×,按更高倍率吃 credits:GPT-5.6 / 5.5 为标准的 2.5×,GPT-5.4 为 2×;/fast on|off|status 切,或 config 里 service_tier = "fast"。仅 ChatGPT 登录可用(是 credit 特性);API key 走 API token 价、无 credit 倍率。GPT-5.3-Codex-Spark 是另一个更快、能力较弱的独立模型(研究预览,仅 ChatGPT Pro),有自己的用量限额,和 fast mode 不是一回事。截至 2026-08-05。OpenAI Codex Docs · Speed — Fast mode speeds a supported model up by 1.5x, consuming credits at a higher multiplier: 2.5x the standard rate for GPT-5.6 / 5.5, 2x for GPT-5.4; toggle with /fast on|off|status or service_tier = "fast" in config. Available only with ChatGPT sign-in (it's a credit feature); an API key uses API token pricing with no credit multiplier. GPT-5.3-Codex-Spark is a separate faster, less-capable model (research preview, ChatGPT Pro only) with its own limits, distinct from fast mode. As of 2026-08-05. This is a ChatGPT-sign-in credit feature; API-key use follows API token pricing without that credit multiplier. Its right use is narrow: on for the crunch hour, off for long automation and CI — the speed wall costs money too, just differently from the capacity wall.最省钱的仍是纪律:管住上下文与输出。prompt 写精、只喂相关文件、AGENTS.md 保持短而准(第 5 章)、子 agent 角色收窄(第 6 章)、闲置 MCP 关掉(第 7 章)、常规活降到 Terra / Luna、别让 agent 复述整份日志。前面几章的每个习惯,现在都能在表里找到分母:上下文纪律,就是 credits 纪律。Discipline still saves most: manage context and output. Tighten prompts, feed only relevant files, keep AGENTS.md short and accurate (chapter 5), narrow child-agent roles (chapter 6), disable idle MCP servers (chapter 7), downshift routine work to Terra / Luna, and don't have the agent recite entire logs. Every earlier habit now has a denominator in the table: context discipline is credit discipline.
— 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。Claude Code(Pro $20 / Max 从 $100,均含在订阅里)的主场在终端里的编排:计划、记忆、subagents、hooks 拼成的那套「派活 + 验收」循环,是它写得最透的地方。4注 4Note 4Claude · 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 均含 Claude Code。截至 2026-08-05。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 include Claude Code. As of 2026-08-05.Cursor(Pro $20 起,顶上还有 Pro+ / Ultra)的主场在编辑器:你要「陪着改、盯每一步 diff」的工作节奏时,它的 agent 工作台仍是最顺的入口。5注 5Note 5Cursor · Pricing —— Hobby 免费;个人付费从 Pro $20/月起(页面另有 Pro+ / Ultra 更高档);Teams $40/座·月,Enterprise 定制。卖点为 Agent 请求、code review、cloud agents。截至 2026-08-05。Cursor · Pricing — Hobby free; paid individual from Pro at $20/mo (the page also has higher Pro+ / Ultra tiers); Teams $40/seat·mo, Enterprise custom. The pitch is Agent requests, code review, and cloud agents. As of 2026-08-05.Copilot 的主场在 GitHub 本体:Pro $10 起能把 issue 派给 Copilot 自家的 cloud agent、收回一个 PR;若要再转派给 Claude、Codex 这些第三方 coding agent,则要 Pro+ $39 或 Max。你的团队整个活在 GitHub 流程里时,它是那条流程自带的手。6注 6Note 6GitHub Copilot · Plans —— Free / Pro $10 / Pro+ $39 / Max $100(每人·月),另有 Business 与 Enterprise。Copilot 自家的 cloud agent /「把活派给 Copilot,它创建一个 pull request」从 Pro 起;但对比表里「把任务转派给 Claude、OpenAI Codex 等第三方 coding agent(Preview)」明确是 Pro 不含、Pro+ 与 Max 才含。页面顶部 Pro 摘要与详细对比行表述冲突,这里按更具体的 feature row 写 Pro+ 起。截至 2026-08-05。GitHub Copilot · Plans — Free / Pro $10 / Pro+ $39 / Max $100 (per user·mo), plus Business and Enterprise. Copilot's own cloud agent / 'assign work to Copilot and it creates a pull request' starts at Pro; but the comparison row for 'delegate tasks to third-party coding agents' such as Claude and OpenAI Codex (Preview) explicitly marks Pro not included and Pro+ / Max included. The page's top-level Pro summary conflicts with that detailed row, so this book follows the more specific feature row: Pro+ and above. As of 2026-08-05.Warp 自我定位「面向 agentic coding 的现代终端」(Build $20/月):活以终端会话本身为中心时,它贴得更近。7注 7Note 7Warp · Pricing —— 自我定位「面向 agentic coding 的现代终端」;Free / Build $20 / Max $200(月付),另有 Business $50/人·月(年付 $45)与 Enterprise。截至 2026-08-05。Warp · Pricing — self-positioned as 'a modern terminal for agentic coding'; Free / Build $20 / Max $200 monthly, plus Business $50/user·mo ($45 annual) and Enterprise. As of 2026-08-05.Now the division — no hit piece, just home turf. Claude Code (Pro $20 / Max from $100, both bundled in the subscription) owns orchestration in the terminal: the delegate-and-verify loop built from plan, memory, subagents, and hooks is where it's written most thoroughly.4注 4Note 4Claude · 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 均含 Claude Code。截至 2026-08-05。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 include Claude Code. As of 2026-08-05.Cursor (from Pro $20, with Pro+ / Ultra above) owns the editor: when your rhythm is 'work alongside, watch every diff,' its agent workbench remains the smoothest entrance.5注 5Note 5Cursor · Pricing —— Hobby 免费;个人付费从 Pro $20/月起(页面另有 Pro+ / Ultra 更高档);Teams $40/座·月,Enterprise 定制。卖点为 Agent 请求、code review、cloud agents。截至 2026-08-05。Cursor · Pricing — Hobby free; paid individual from Pro at $20/mo (the page also has higher Pro+ / Ultra tiers); Teams $40/seat·mo, Enterprise custom. The pitch is Agent requests, code review, and cloud agents. As of 2026-08-05.Copilot owns GitHub itself: from Pro at $10, assign an issue to Copilot's own cloud agent and collect a PR; delegating onward to third-party coding agents such as Claude and Codex requires Pro+ at $39 or Max. When your team lives entirely in GitHub flow, it's the hand that flow comes with.6注 6Note 6GitHub Copilot · Plans —— Free / Pro $10 / Pro+ $39 / Max $100(每人·月),另有 Business 与 Enterprise。Copilot 自家的 cloud agent /「把活派给 Copilot,它创建一个 pull request」从 Pro 起;但对比表里「把任务转派给 Claude、OpenAI Codex 等第三方 coding agent(Preview)」明确是 Pro 不含、Pro+ 与 Max 才含。页面顶部 Pro 摘要与详细对比行表述冲突,这里按更具体的 feature row 写 Pro+ 起。截至 2026-08-05。GitHub Copilot · Plans — Free / Pro $10 / Pro+ $39 / Max $100 (per user·mo), plus Business and Enterprise. Copilot's own cloud agent / 'assign work to Copilot and it creates a pull request' starts at Pro; but the comparison row for 'delegate tasks to third-party coding agents' such as Claude and OpenAI Codex (Preview) explicitly marks Pro not included and Pro+ / Max included. The page's top-level Pro summary conflicts with that detailed row, so this book follows the more specific feature row: Pro+ and above. As of 2026-08-05.Warp positions itself as 'a modern terminal for agentic coding' (Build $20/mo): when work centers on the terminal session itself, it sits closer.7注 7Note 7Warp · Pricing —— 自我定位「面向 agentic coding 的现代终端」;Free / Build $20 / Max $200(月付),另有 Business $50/人·月(年付 $45)与 Enterprise。截至 2026-08-05。Warp · Pricing — self-positioned as 'a modern terminal for agentic coding'; Free / Build $20 / Max $200 monthly, plus Business $50/user·mo ($45 annual) and Enterprise. As of 2026-08-05.Codex 自己的主场,是这本书的整条主线:本地分工 + 隔离执行 + GitHub 验收 —— 已经付着 ChatGPT、常做「subagents 分头查、独立任务上云、回来收 PR」的人,先用的是计划内含用量;到限后才面对 credits / 升档的选择,不能再说成永远「不额外付费」。1注 1Note 1OpenAI Codex 文档 · Pricing —— 页面把 Work 与 Codex 列为 ChatGPT Free、Go、Plus、Pro、Business、Enterprise、Edu 的内含能力,共用计划用量;Free 是 quick、Go 是 light、Plus 起更完整。价格:Free $0、Go $8、Plus $20、Pro 从 $100(选比 Plus 多 5× 或 20× 用量)、Business $20/人·月(年付,月付 $25)。实际 Work 入口仍受平台、账号与 rollout 限制;API key 旁路按 API token 计费、不含 Codex cloud。截至 2026-08-05。OpenAI Codex Docs · Pricing — the page lists Work and Codex as included capabilities across ChatGPT Free, Go, Plus, Pro, Business, Enterprise, and Edu, sharing plan usage; Free is quick, Go light, and Plus upward more complete. Prices: Free $0, Go $8, Plus $20, Pro from $100 (choose 5x or 20x more usage than Plus), Business $20/user·mo annual ($25 monthly). Actual Work access still depends on platform, account, and rollout; the API-key side road uses API token billing and excludes Codex cloud. As of 2026-08-05.9注 9Note 9OpenAI Help Center · Credits in ChatGPT —— 先用计划内含额度;eligible Plus / Pro 到限后可加购 credits,不用立刻升级。Work 与 Codex 在可用处共用计划用量和 credits;购买后 12 个月过期、不滚存、不可退款。实际可购性和适用功能因计划 / 地区 / rollout 而异。截至 2026-08-05。OpenAI Help Center · Credits in ChatGPT — included plan usage is consumed first; eligible Plus / Pro users can buy credits after reaching the limit instead of immediately upgrading. Where available, Work and Codex share plan usage and credits; purchases expire after 12 months, do not roll over, and are non-refundable. Purchase eligibility and supported features vary by plan / region / rollout. As of 2026-08-05.组合原则仍只一句:入口跟着活走,订阅别叠着买 —— 每加一份订阅,都该对应一类你真在做、且 Codex 明显干得不如它的活。Codex's own home turf is this book's through-line: local delegation + isolated execution + GitHub review — for someone already paying for ChatGPT who often has subagents investigate, independent tasks run in cloud, and PRs come back, included plan usage comes first; only after the limit do credits / upgrades enter, so it can't honestly be called 'never extra cost.'1注 1Note 1OpenAI Codex 文档 · Pricing —— 页面把 Work 与 Codex 列为 ChatGPT Free、Go、Plus、Pro、Business、Enterprise、Edu 的内含能力,共用计划用量;Free 是 quick、Go 是 light、Plus 起更完整。价格:Free $0、Go $8、Plus $20、Pro 从 $100(选比 Plus 多 5× 或 20× 用量)、Business $20/人·月(年付,月付 $25)。实际 Work 入口仍受平台、账号与 rollout 限制;API key 旁路按 API token 计费、不含 Codex cloud。截至 2026-08-05。OpenAI Codex Docs · Pricing — the page lists Work and Codex as included capabilities across ChatGPT Free, Go, Plus, Pro, Business, Enterprise, and Edu, sharing plan usage; Free is quick, Go light, and Plus upward more complete. Prices: Free $0, Go $8, Plus $20, Pro from $100 (choose 5x or 20x more usage than Plus), Business $20/user·mo annual ($25 monthly). Actual Work access still depends on platform, account, and rollout; the API-key side road uses API token billing and excludes Codex cloud. As of 2026-08-05.9注 9Note 9OpenAI Help Center · Credits in ChatGPT —— 先用计划内含额度;eligible Plus / Pro 到限后可加购 credits,不用立刻升级。Work 与 Codex 在可用处共用计划用量和 credits;购买后 12 个月过期、不滚存、不可退款。实际可购性和适用功能因计划 / 地区 / rollout 而异。截至 2026-08-05。OpenAI Help Center · Credits in ChatGPT — included plan usage is consumed first; eligible Plus / Pro users can buy credits after reaching the limit instead of immediately upgrading. Where available, Work and Codex share plan usage and credits; purchases expire after 12 months, do not roll over, and are non-refundable. Purchase eligibility and supported features vary by plan / region / rollout. As of 2026-08-05. The combination principle remains one line: entrances follow the work; don't stack subscriptions — each added subscription should map to work you actually do and that Codex demonstrably handles worse.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:计划内用量见底、模型吃力、还是嫌慢 —— 别再记消息条数,记任务类型与模型。For one week, log each can't-use moment: included usage dry, model straining, or too slow — don't count messages; record task type and model.
把批量读 / 扫降到 Luna、常规实现放 Terra、难活留 Sol,跑一周对照 input / cached input / output 的 credit 分布。Downshift batched reading / scanning to Luna, routine implementation to Terra, and keep hard work on Sol; compare the input / cached-input / output credit mix after a week.
只在赶工的一小时开 /fast,完事就关 —— 对照 standard 基线,确认快约 50% 是否值多 150% credits。Use /fast only for the crunch hour, then turn it off — against the standard baseline, decide whether roughly 50% more speed was worth 150% more credits.
入口跟着活走,
订阅别叠着买.
Entrances follow the work;
don't stack subscriptions.
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
