diff --git a/src/content/guides-en/codex-plan-value-analysis.md b/src/content/guides-en/codex-plan-value-analysis.md new file mode 100644 index 0000000..aaf798e --- /dev/null +++ b/src/content/guides-en/codex-plan-value-analysis.md @@ -0,0 +1,95 @@ +--- +title: "Codex Plan Value Analysis (2026): Plus, Pro 5x, or Pro 20x?" +description: "OpenAI Codex is now subscription-based: Plus $20/month, Pro 5x $100/month, Pro 20x $200/month. This review runs the numbers with official pricing verified 2026-08-19: per-model 5-hour window message ranges for GPT-5.6 Sol/Terra/Luna, the credits system, and what to do when you hit limits - so you can choose between subscriptions, pay-as-you-go API, and competing plans like Claude Max or domestic coding plans." +date: "2026-08-19" +article_type: review +tags: [codex, chatgpt, openai, pricing, subscription] +pillar: plans +content_status: keep +locale_strategy: mirrored +draft: false +--- + +## Bottom Line + +- **Heavy Codex users** (2+ hours of agent sessions in the CLI/IDE daily): Pro 5x ($100/mo) is the sweet spot - 5x Plus local-message capacity and the full GPT-5.6 family. +- **All-day multi-agent workloads that regularly hit weekly caps**: Pro 20x ($200/mo) is the only tier that makes sense; otherwise Pro 5x plus occasional credits is cheaper. +- **Occasional use**: Plus ($20/mo) with Luna/Terra is plenty - don't jump to Pro early. +- **Very spiky usage or team procurement**: run the API math first; teams should look at Business directly. +- Same conclusion as yesterday's [Claude Max analysis](/en/guides/claude-max-plan-value-analysis/): these subscriptions buy "ceiling plus convenience"; for users outside the US there are also payment and network frictions, so budget-sensitive buyers should weigh domestic coding plans. + +--- + +## Plan Structure (verified against official docs 2026-08-19) + +Codex now lives inside the ChatGPT subscription system, with five paid options: + +| Tier | Monthly | Position | +|------|------|------| +| Plus | $20 | Entry - Codex CLI / IDE / App included | +| Pro 5x | $100 | Heavy users - 5x Plus local-message capacity | +| Pro 20x | $200 | Top individual tier - 20x capacity + unlimited Voice | +| Business | per seat | Team management, automated Code Review, Slack/Linear integration | +| API Key | usage-based | Flexible, token-billed, cost scales with usage | + +Pro tiers add: higher output, priority access at peak times, the GPT-5.3-Codex-Spark research preview (Pro only), and more Voice time (unlimited on Pro 20x). + +## How Usage Works: Per-Model 5-Hour Windows + +Codex usage is not a fixed message count - it floats with model and task complexity, with a rolling 5-hour window plus possible weekly caps, shared between local sessions and cloud. Official **local messages / 5-hour window** by tier: + +| Model | Plus | Pro 5x | Pro 20x | +|------|------|--------|---------| +| GPT-5.6 Sol (hardest reasoning) | 10-100 | 50-500 | 200-2,000 | +| GPT-5.6 Terra (everyday workhorse) | 25-200 | 125-1,000 | 500-4,000 | +| GPT-5.6 Luna (high volume) | 250-2,000 | 1,250-10,000 | 5,000-40,000 | +| GPT-5.5 | 15-80 | 75-400 | 300-1,600 | +| GPT-5.4 | 20-100 | 100-500 | 400-2,000 | +| GPT-5.4 mini | 60-350 | 300-1,750 | 1,200-7,000 | + +Key takeaways: + +1. **Model choice matters more than tier choice**: within the same tier, Luna's allowance is 20-25x Sol's. Use Luna/Terra for routine work and save Sol for genuinely hard problems - the 5-hour window will last much longer. +2. **Business seats = Plus-level usage**, but with team management; Enterprise/Edu with flexible pricing scale by credits with no fixed rate limits. +3. **When you hit a limit**, you can buy credits to keep working, switch to a smaller model, or (all users) run extra local chats with an API key at standard API rates. + +## Credits Pricing: GPT-5.6 Averages 5-40 Credits per Message + +Credits are token-billed. Official rate card (credits / 1M tokens, input/cached/output): + +| Model | Input | Cached input | Output | +|------|------|---------|------| +| GPT-5.6 Sol | 125 | 12.5 | 750 | +| GPT-5.6 Terra | 50 | 5 | 300 | +| GPT-5.6 Luna | 5 | 0.5 | 30 | +| GPT-5.5 | 125 | 12.5 | 750 | +| GPT-5.4 | 62.5 | 6.25 | 375 | +| GPT-5.4 mini | 18.75 | 1.875 | 113 | + +OpenAI states GPT-5.6 averages 5-40 credits per message. Rough math: a long-context agent task (hundreds of thousands of input tokens) costs roughly 15-25 credits on Terra, 40-90 on Sol. If you buy credits every week, your tier is too low - upgrading to Pro is usually cheaper than sustained credit purchases. + +## Running the Numbers: Plus / Pro 5x / Pro 20x / API + +- **Plus $20**: for under ~50 medium sessions a month, mostly Luna/Terra. At Terra's 25-200 messages / 5h, 1-2 focused hours a day is generally fine. +- **Pro 5x $100**: Terra 125-1,000 / 5h, Sol 50-500 - the sweet spot for professional developers. Versus API (Sol output at 750 credits / 1M tokens), 4-6 Sol-class sessions a day already justifies the flat fee. +- **Pro 20x $200**: near-unlimited-feeling quota for all-day agents, parallel tasks, and unlimited Voice. If your 5x weekly cap keeps hitting, the $100 increment is likely cheaper than rate-limit waiting plus credit purchases. +- **API Key**: more flexible when usage is spiky or you want pure token billing; downsides are unpredictable cost and no bundled Voice/Web/cloud features. + +## Compared with Claude Max and Domestic Plans + +- **Claude Max 5x/20x ($100/$200)**: nearly identical structure (5-hour window + weekly caps + shared pool + credits). The difference is the model ecosystem: Claude's Opus family still tops this site's coding score (9.6/10) for long code generation, while Codex's GPT-5.6 family is stronger at agent orchestration and native GitHub integration (@Codex PR/issue delegation, automated code review). If you're torn, see the [Claude Code vs Codex comparison](/en/compare/claude-code-vs-codex/). +- **Domestic coding plans (GLM / Volcengine Ark / Bailian)**: typically ¥50-200/month, payable in RMB with direct domestic access. Codex requires chatgpt.com access and international payment, which has real friction in mainland China (the site's codex-cli data rates china_friendly 1/10). Budget-sensitive users in China should start with domestic plans - see our [GLM-5.3 Coding Plan review](/en/guides/glm-5.3-coding-plan-review/) and [GLM vs DeepSeek vs Kimi comparison](/en/compare/glm-vs-deepseek-vs-kimi-2026/). + +## Who Should Buy + +**Plus**: light usage, Luna/Terra is enough, budget-sensitive. +**Pro 5x**: Codex is your primary dev tool, 2h+ daily, need Sol-class reasoning, or you keep hitting Plus limits. +**Pro 20x**: all-day agent workflows, heavy Voice user, 5x weekly cap frequently hit. +**Business**: teams needing automated code review, Slack/Linear integration, and compliance controls. +**API Key**: spiky usage, per-token cost control, or you already run your own agent framework. + +## Sources and Review Dates + +- Plan structure, per-tier usage tables, credits rate card, Voice allowances, feature matrix: OpenAI Codex official docs, [Pricing / Usage limits](https://developers.openai.com/codex/pricing), verified 2026-08-19. +- Usage tables are official **ranges** (they float with task complexity), not fixed promises; both OpenAI and Anthropic reserve the right to adjust quotas and pricing - always defer to the official pages before purchasing. +- Re-verify pricing quarterly; if you spot stale numbers, please open an issue on the GitHub repo. diff --git a/src/content/guides-en/deepseek-harness-intro.md b/src/content/guides-en/deepseek-harness-intro.md index 117e3cb..1c3014d 100644 --- a/src/content/guides-en/deepseek-harness-intro.md +++ b/src/content/guides-en/deepseek-harness-intro.md @@ -114,6 +114,7 @@ In addition to the Web UI and CLI, DeepSeek Harness offers a Python SDK for inte DeepSeek Harness plugins are discoverable through the GitHub [`dsh-plugin` topic](https://github.com/topics/dsh-plugin). You can: - Browse existing community plugins +- Browse installable plugins in the community registry [dsplugin.app (dsh plugin registry)](https://dsplugin.app/), filterable by source and manifest status (community-maintained, not official) - Tag your own plugin with `dsh-plugin` to make it discoverable - Follow the official [plugin development guide](https://github.com/deepseek-ai/deepseek-harness/tree/master/docs) to write your own diff --git a/src/content/guides/codex-plan-value-analysis.md b/src/content/guides/codex-plan-value-analysis.md new file mode 100644 index 0000000..1c1995f --- /dev/null +++ b/src/content/guides/codex-plan-value-analysis.md @@ -0,0 +1,95 @@ +--- +title: "Codex 套餐性价比分析（2026）：Plus、Pro 5x、Pro 20x 怎么选" +description: "OpenAI Codex 现在是订阅制：Plus $20/月、Pro 5x $100/月、Pro 20x $200/月。本文按官方 2026-08-19 价格与用量规则算账：GPT-5.6 Sol/Terra/Luna 各档 5 小时窗口消息数、credits 体系、撞墙后的出路，帮你判断订阅还是 API 按量，以及和 Claude Max、国产 Coding Plan 的横向对比。" +date: "2026-08-19" +article_type: review +tags: [codex, chatgpt, openai, pricing, subscription] +pillar: plans +content_status: keep +locale_strategy: mirrored +draft: false +--- + +## 先说结论 + +- **重度 Codex 用户**（每天在 CLI / IDE 里跑 Agent 会话 2 小时以上）：Pro 5x（$100/月）是甜点区，本地消息量是 Plus 的 5 倍，且能用 GPT-5.6 全系。 +- **全天候多任务、经常撞周上限**：Pro 20x（$200/月）才有意义；一般用量 Pro 5x + 偶尔买 credits 更划算。 +- **只偶尔用**：Plus（$20/月）的 Luna/Terra 档位足够，别急着上 Pro。 +- **用量极不稳定或团队采购**：先算 API 按量账；团队直接看 Business。 +- 和昨天写的 [Claude Max 分析](/zh/guides/claude-max-plan-value-analysis/) 结论一致：这类订阅买的是「天花板 + 省心」，对国内用户还有支付与网络成本，预算敏感者优先看国产 Coding Plan。 + +--- + +## 套餐结构（官方 2026-08-19 核对） + +Codex 现在挂在 ChatGPT 订阅体系下，共 5 种付费方式： + +| 档位 | 月费 | 定位 | +|------|------|------| +| Plus | $20/月 | 入门，Codex CLI / IDE / App 可用 | +| Pro 5x | $100/月 | 重度用户，5 倍于 Plus 的本地消息量 | +| Pro 20x | $200/月 | 最高个人档，20 倍消息量 + 无限制 Voice | +| Business | 按座席 | 团队管理、自动化 Code Review、Slack/Linear 集成 | +| API Key | 按量 | 灵活，token 计费，成本随用量波动 | + +Pro 档额外包含：更高输出、高峰期优先、GPT-5.3-Codex-Spark 研究预览（仅 Pro）、更多 Voice 时长（Pro 20x 无限）。 + +## 用量机制：按模型分档的 5 小时窗口 + +Codex 的用量不是"固定条数"，而是按模型和复杂度浮动，5 小时滚动窗口 + 可能的周上限，本地会话与云端共享。官方给出的各档位**本地消息数 / 5 小时窗口**： + +| 模型 | Plus | Pro 5x | Pro 20x | +|------|------|--------|---------| +| GPT-5.6 Sol（最强推理） | 10-100 | 50-500 | 200-2,000 | +| GPT-5.6 Terra（日常主力） | 25-200 | 125-1,000 | 500-4,000 | +| GPT-5.6 Luna（高吞吐） | 250-2,000 | 1,250-10,000 | 5,000-40,000 | +| GPT-5.5 | 15-80 | 75-400 | 300-1,600 | +| GPT-5.4 | 20-100 | 100-500 | 400-2,000 | +| GPT-5.4 mini | 60-350 | 300-1,750 | 1,200-7,000 | + +几点关键解读： + +1. **选对模型比选对套餐更能省额度**：同一档位下 Luna 的额度是 Sol 的 20-25 倍。日常简单任务用 Luna/Terra，把 Sol 留给真正难的问题，5h 窗口能撑很久。 +2. **Business 座席 = Plus 级用量**，但带团队管理；Enterprise/Edu 灵活定价则按 credits 走，无固定限流。 +3. **撞墙后可买 credits 续跑**，或临时切小模型；所有用户还可以用 API Key 跑额外本地会话（按标准 API 费率）。 + +## credits 计价：GPT-5.6 平均每条 5-40 credits + +撞墙后的 credits 按 token 计费，官方费率（credits / 每 1M tokens，输入/缓存/输出）： + +| 模型 | 输入 | 缓存输入 | 输出 | +|------|------|---------|------| +| GPT-5.6 Sol | 125 | 12.5 | 750 | +| GPT-5.6 Terra | 50 | 5 | 300 | +| GPT-5.6 Luna | 5 | 0.5 | 30 | +| GPT-5.5 | 125 | 12.5 | 750 | +| GPT-5.4 | 62.5 | 6.25 | 375 | +| GPT-5.4 mini | 18.75 | 1.875 | 113 | + +官方说明 GPT-5.6 平均每条消息消耗 5-40 credits。粗算：一次长上下文 Agent 任务（几十万输入 token）在 Terra 上约消耗 15-25 credits，Sol 上约 40-90 credits。如果你每周都要买 credits，说明套餐档位低了——升级 Pro 通常比持续买 credits 便宜。 + +## 算账：Plus / Pro 5x / Pro 20x / API + +- **Plus $20**：适合每月 <50 次中等会话、且以 Luna/Terra 为主的人。按 Terra 25-200 条 / 5h 算，一天认真用 1-2 小时基本够。 +- **Pro 5x $100**：Terra 125-1,000 条 / 5h，Sol 50-500 条——重度开发者的甜点区。对比 API 按量（Sol 750 输出 credits / 1M tokens，折合订阅外购买价），每天 4-6 次 Sol 级会话就值回票价。 +- **Pro 20x $200**：接近"无感额度"，适合整天挂 Agent、多任务并行、Voice 无限时长。如果你的 5x 周上限经常触顶，20x 的增量成本（$100）大概率低于限流等待和 credits 采购。 +- **API Key**：用量波动大、或想完全按 token 付费时更灵活；缺点是成本不可控、没有订阅附带的 Voice/Web/云功能。 + +## 和 Claude Max / 国产方案对比 + +- **Claude Max 5x/20x（$100/$200）**：结构几乎一模一样（5 小时窗口 + 周上限 + 全端共享 + credits 续跑）。差异在模型生态：Claude 的 Opus 系在长代码生成上仍是站内评分最高（9.6/10），Codex 的 GPT-5.6 系在 Agent 编排和 GitHub 原生集成（@Codex 委派 PR/issue、自动 Code Review）上更强。纠结选型的看 [Claude Code vs Codex 对比](/zh/compare/claude-code-vs-codex/)。 +- **国产 Coding Plan（GLM / 方舟 / 百炼）**：月费 ¥50-200，人民币直付、国内直连。Codex 的 chatgpt.com 访问与支付在国内有现实摩擦（站内 `codex-cli` 数据 china_friendly 仅 1/10）。预算敏感、人在国内：优先国产方案，详见 [GLM-5.3 Coding Plan 评测](/zh/guides/glm-5.3-coding-plan-review/) 和 [GLM vs DeepSeek vs Kimi 对比](/zh/compare/glm-vs-deepseek-vs-kimi-2026/)。 + +## 适合谁买 + +**Plus**：轻度使用、Luna/Terra 够用、预算敏感。 +**Pro 5x**：主力开发工具是 Codex、每天 2h+、需要 Sol 档推理、或常被 Plus 限流。 +**Pro 20x**：全天 Agent 工作流、Voice 重度用户、5x 周上限频繁触顶。 +**Business**：团队共享、需要自动化 Code Review 和 Slack/Linear 集成、合规审计。 +**API Key**：用量波动大、团队按量管控、或已跑在自有 Agent 框架里。 + +## 数据来源与复核 + +- 套餐结构、各档位用量表、credits 费率、Voice 时长、功能矩阵：OpenAI Codex 官方文档 [Pricing / Usage limits](https://developers.openai.com/codex/pricing)，2026-08-19 核对。 +- 用量表为官方给出的**范围值**（随任务复杂度浮动），非固定承诺；Anthropic/OpenAI 均保留调整额度的权利，购买前请以官方页面为准。 +- 建议每季度复核一次；发现过期数字欢迎在 GitHub 仓库提 issue。 diff --git a/src/content/guides/deepseek-harness-intro.md b/src/content/guides/deepseek-harness-intro.md index 0283714..5e3c347 100644 --- a/src/content/guides/deepseek-harness-intro.md +++ b/src/content/guides/deepseek-harness-intro.md @@ -114,6 +114,7 @@ pnpm dsh web DeepSeek Harness 的插件通过 GitHub 的 [`dsh-plugin` topic](https://github.com/topics/dsh-plugin) 来发现。你可以： - 浏览社区已有的插件 +- 在社区目录 [dsplugin.app（dsh plugin registry）](https://dsplugin.app/) 按来源与 Manifest 状态筛选可安装插件（社区维护，非官方） - 把自己的插件打上 `dsh-plugin` topic 让它被搜索到 - 参考官方文档 [插件开发指南](https://github.com/deepseek-ai/deepseek-harness/tree/master/docs) 写自己的插件