Add retirement_date field and backfill it across all first-party labs

[Tangerg]

Summary

Implements the feature proposed in #1921: there was no way to record when a model stops being served by the provider — only that it is deprecated.

Why a separate retirement_date? Isn't status = "deprecated" enough?

No — they describe two different things, and providers treat them as two distinct events:

• Deprecated = announced, no longer recommended, but still callable.
• Retired / shutdown = actually removed; requests after this date fail.

Every major lab publishes both a deprecation announcement and a later shutdown date, with a deliberate migration buffer in between. For example, OpenAI's 2026-04-22 notice:

Deprecated (announced)	Shutdown date	Model
2026-04-22	2026-07-23	computer-use-preview → gpt-5.4-mini

That ~3-month gap is the window developers get to migrate. status = "deprecated" alone tells you a model is on its way out but not when it actually disappears — which is exactly the information consumers need to plan migrations, warn users, or hide soon-dead models.

So this PR keeps the two orthogonal:

• status = "deprecated" → the lifecycle state (announced, still served).
• retirement_date → the date it is (or was) shut off.

A model can be deprecated today with a future retirement_date (still usable, but scheduled to die), e.g. gpt-3.5-turbo → 2026-10-23. No new status value is introduced; the existing deprecated is reused. OpenAI calls this the "shutdown date" and Anthropic the "retirement date"; retirement_date was chosen as the neutral name.

What's in this PR

1. Schema (packages/core/src/schema.ts) — adds an optional retirement_date field, same ^\d{4}-\d{2}(-\d{2})?$ format as release_date / last_updated.
2. Docs — updates the field reference + example TOML in README.md and the field table in AGENTS.md.
3. Data — backfills 49 canonical first-party models across every lab in the repo that publishes an official retirement/deprecation schedule.

The field flows through to api.json automatically, and all mirror/aggregator providers (Vertex, Bedrock, OpenRouter, Vercel, Azure, …) inherit it via symlinks and extends.

Convention

• retirement_date is set whenever the provider publishes an official date.
• status = "deprecated" is set only when the model is already retired or clearly superseded. Still-active models that merely have a future published date get retirement_date without a status change.

Coverage by provider (official sources only)

Provider	Models updated	Source
OpenAI	6 (gpt-3.5-turbo, gpt-4, gpt-4-turbo, gpt-4o-2024-05-13, o1-preview, o1-mini)	https://developers.openai.com/api/docs/deprecations
Anthropic	9 (claude-3*, claude-3-5*, claude-3-7-sonnet, claude-sonnet-4-20250514, claude-opus-4-20250514)	https://platform.claude.com/docs/en/about-claude/model-deprecations
Google (Gemini)	11	https://ai.google.dev/gemini-api/docs/deprecations
Mistral	14 legacy models	https://docs.mistral.ai/getting-started/models/
Cohere	2 (c4ai-aya-expanse-8b, c4ai-aya-vision-8b → 2026-04-04)	https://docs.cohere.com/docs/deprecations
DeepSeek	2 (deepseek-chat, deepseek-reasoner → 2026-07-24)	https://api-docs.deepseek.com/updates
Moonshot / Kimi	5 (kimi-k2-* → 2026-05-25)	https://platform.kimi.ai/docs/models

Checked, no official retirement schedule for tracked models (left unchanged)

xAI (Grok), Perplexity (Sonar), Alibaba (Qwen / Model Studio), Amazon Bedrock, NVIDIA, Zhipu (GLM), StepFun, MiniMax, Upstage, Sarvam. For these, the provider either publishes no formal schedule, or the only dated retirements concern model IDs not tracked in this repo.

Notes for reviewers

• Only canonical first-party model files were edited; mirrors inherit automatically.
• Google publishes shutdown dates explicitly framed as the earliest possible retirement; they're included as the official published value.
• Partner platforms (Amazon Bedrock, Vertex AI) set their own schedules per Anthropic's docs; the symlinked mirrors currently inherit the canonical dates, consistent with how cost/limits are already mirrored. Happy to special-case if maintainers prefer.
• Field name retirement_date is open to bikeshedding.

Validation

bun run validate passes; retirement_date is present in the generated output for the affected models and their mirrors.

[LaOzhOy1]

Strong +1. Keeping retirement_date orthogonal to status = "deprecated" is the key insight here — "announced but still callable" and "actually shut off" are genuinely different states, and only the latter is a hard deadline. The 3-month OpenAI example in the description makes the case well. Backfilling from official provider pages (rather than guessing) is exactly the right level of rigor for this repo.