LLM模型弃用通知中位数为183天，最短仅20天
The median model deprecation notice is 183 days. The shortest this year was 20.
做LLM工程的同学必看，这篇直接给出了扫描代码中模型标识和参数、对接CI流水线的具体做法，能帮你提前应对厂商突如其来的弃用通知。
Disclosure up front: I maintain the open source tool linked at the end. It is MIT, there is no paid tier and nothing to sign up for. The numbers come first and every one of them is checkable against the sources at the bottom.
提前声明：我维护着文末链接的开源工具。它采用 MIT 许可证，没有付费层级，也无需注册。数据优先，且每一项均可对照底部的来源进行核查。
I run an LLM product in production, so I transcribed both providers' deprecation pages into machine-readable JSON to find out what my own code was exposed to. 116 model entries, 33 from Anthropic and 83 from OpenAI, as the pages stood on 2026-09-20. What fell out:
我在生产环境中运行一个 LLM 产品，因此我将两家提供商的弃用页面转录为机器可读的 JSON，以查明我的代码暴露于哪些情况。截至 2026-09-20，共有 116 个模型条目，其中 Anthropic 33 个，OpenAI 83 个。结果如下：
- 43 entries still have a retirement date ahead of them. The nearest is the sora-2 family, 2 days from today.
- Across the 102 entries carrying both an announcement and a retirement date, the median gap is 183 days. The shortest is 20, for gpt-5.4-cyber.
- In 2026 alone, OpenAI published 8 separate announcement dates and Anthropic 4.
- 仍有 43 个条目具有未来的退役日期。最近的是 sora-2 系列，距离今天仅 2 天。
- 在同时包含公告日期和退役日期的 102 个条目中，中位间隔为 183 天。最短的是 20 天，对应 gpt-5.4-cyber。
- 仅在 2026 年，OpenAI 就发布了 8 个独立的公告日期，Anthropic 则为 4 个。
The 183 against the 20 is the part worth sitting with. Six months of notice is generous enough that a team learns the notice does not matter, and then a 20 day one arrives in a week nobody planned for.
183 天与 20 天的对比是值得我们深思的部分。六个月的通知期足够慷慨，以至于团队会习以为常，认为通知无关紧要；然而，一个仅有 20 天的通知会在一个毫无准备的时间点突然到来。
Retirement is the loud failure. The quiet one is a parameter. Per Anthropic's page, temperature, top_p and top_k return a 400 on Claude Opus 4.7 and later, and the Python SDK from v1.0 removes them, so passing them raises a TypeError. Nothing in a test suite catches that if the tests never hit the new model, and no changelog files it under "deprecation". If your reproducibility depends on temperature=0, as mine does, that is not a small migration.
退役是显性的失败。隐性的失败则是一个参数。根据 Anthropic 的页面说明，temperature、top_p 和 top_k 在 Claude Opus 4.7 及更高版本中会返回 400 错误，而 Python SDK 从 v1.0 起移除了这些参数，因此传入它们会引发 TypeError。如果测试套件从未触及新模型，就不会捕获到这一点；而且变更日志也不会将其归类为“弃用”。如果你的可复现性依赖于 temperature=0（正如我所依赖的那样），那将不是一次简单的迁移。
What I do about it now: a CLI that scans a tree for model identifiers, crosses them with those dates, and exits with a code CI can act on, so the build starts failing the day a provider gives notice rather than the day the model dies. Deprecated parameters are located with the stdlib ast module, so it finds the keyword argument itself and not a string that looks like one. Given a usage export, each finding also carries its share of real traffic, which separates a dead constant from 40% of production.
我现在采取的措施：一个 CLI 工具，用于扫描树状结构中的模型标识符，并与这些日期进行交叉比对，然后退出一个 CI 可以据此采取行动的状态码，以便在提供商发出通知的那一天构建就开始失败，而不是等到模型真正死亡的那一天。弃用的参数通过标准库的 ast 模块定位，因此它能找到实际的关键词参数本身，而非看起来像参数的字符串。给定一个使用量导出文件，每个发现还附带其对应的真实流量占比，从而区分出死常量与占生产环境 40% 的活跃部分。
uvx depdesk check .uvx depdesk check .Standard library only, no dependencies, Python 3.9 and up, MIT: https://github.com/Ort0x36/deprecation-desk
仅依赖标准库，无外部依赖，支持 Python 3.9 及以上版本，MIT 协议：https://github.com/Ort0x36/deprecation-desk
The catalog is hand-transcribed JSON carrying the date it was verified, and a scheduled job checks whether the provider pages moved since. I decided against scraping: a scraper that half works produces a catalog that looks maintained and is not, which is the exact failure this is supposed to prevent.
该目录是手动转录的 JSON，包含验证日期，并且有一个计划任务会检查提供商页面自上次以来是否发生过变动。我决定不进行网页抓取：一个半吊子的抓取器会产生一个看似维护良好实则不然的目录，而这正是本工具旨在防止的故障类型。
The question I actually want answered here: which provider should come after these two? Bedrock and Azure OpenAI are the two I hear most, and they are a different problem, because the retirement dates belong to the platform rather than to the lab.
我真正想在这里得到回答的问题是：在这两者之后，哪家提供商应该排在后面？Bedrock 和 Azure OpenAI 是我听到最多的两个，它们属于不同的问题，因为退役日期属于平台而非实验室。
Sources
来源
- Anthropic model deprecations: https://platform.claude.com/docs/en/about-claude/model-deprecations
- OpenAI deprecations: https://developers.openai.com/api/docs/deprecations
- The transcription those counts come from, with its verification date: https://github.com/Ort0x36/deprecation-desk/blob/main/depdesk/data/catalog.json
- Anthropic 模型弃用：https://platform.claude.com/docs/en/about-claude/model-deprecations
- OpenAI 弃用：https://developers.openai.com/api/docs/deprecations
- 这些计数所依据的转录数据及其验证日期：https://github.com/Ort0x36/deprecation-desk/blob/main/depdesk/data/catalog.json
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力