Skip to content
← Back to blog
Engineering·August 16, 2026·6 min read

OpenAI gives six months' notice before retiring a model. Anthropic gives sixty days.

OpenAI's own policy promises six months before a model retires. Anthropic's promises sixty days. What that gap costs a team that pinned a model ID.

OpenAI's own deprecation page states a floor of six months' notice before it retires a generally available model. Anthropic's states a floor of sixty days, about a third of that. Both companies mostly deliver right at their own floor rather than comfortably above it, which means the real planning question for a team building on either API isn't whether a model will eventually get retired. It's whether your migration timeline assumes the generous number or the one the vendor actually gives.

What "deprecated" actually means

Anthropic's documentation defines four states for a model: active (fully supported), legacy (no more updates, deprecation possible later), deprecated (still functional but not recommended, with a replacement and a retirement date assigned) and retired (no longer available; requests fail). OpenAI's page works the same way in practice even without naming the stages: an announcement date, a shutdown date, and a named replacement. The distinction that catches teams off guard is between deprecated and retired. A deprecated model still answers your API calls. A retired one returns an error, on every request, starting the day the vendor named months earlier.

The two policies, in the vendors' own words

OpenAI states its policy directly on its deprecations page: "Unless safety or compliance concerns require a faster timeline, we provide the following minimum notice periods before model retirement: Generally available models: At least 6 months. Specialized variants of generally available models: At least 3 months... Preview models... may be retired with much shorter notice, such as 2 weeks." Anthropic's model-deprecations page states its own floor more simply: "Anthropic notifies customers with active deployments for models with upcoming retirements, providing at least 60 days' notice before model retirement for publicly released models."

VendorGA model floorSpecialized variant floorShortest published tier
OpenAIAt least 6 monthsAt least 3 months (chat, Codex, deep-research variants)Preview models, as little as 2 weeks
AnthropicAt least 60 daysSame floor applies; no separate variant tier published60 days is the stated minimum across the board

What the actual track record shows

Policy pages state a floor, not a promise of more, and the last year of actual retirements on both sides sits close to that floor rather than above it. Anthropic's four most recent 2026 retirements: Claude Opus 4.1 (announced June 5, retired August 5, 61 days), Claude Sonnet 4 and Opus 4 (announced April 14, retired June 15, 62 days), Claude Haiku 3 (announced February 19, retired April 20, 60 days) and Claude Haiku 3.5 (announced December 19, 2025, retired February 19, 2026, 62 days). All four land within two days of the stated 60-day minimum. That's tighter than Anthropic's own history: Claude Sonnet 3.7 got 114 days in late 2025, and Claude Opus 3 got 189 days in mid-2025. The floor hasn't moved, but recent practice has moved toward it.

OpenAI's 2026 generally available retirements sit at their own floor in the same way. The legacy GPT snapshots retired in this batch (including gpt-3.5-turbo-0125 and gpt-4-0613) were announced April 22 and shut down October 23, 184 days. The GPT-5 and o3 snapshot family was announced June 11 and shut down December 11, 183 days. Both are essentially exactly six months, not a generous margin past it.

The specialized-variant trap

The gap that actually catches a team is the specialized-variant tier, because the model IDs in it don't look temporary. gpt-5.2-chat-latest and gpt-5.3-chat-latest were announced May 8, 2026 and shut down August 10, 2026, 94 days, right at OpenAI's 3-month floor for chat, Codex and deep-research variants. A "-latest" alias reads like a moving target that always points at the current model, which is exactly why teams pin it for convenience. What it actually is, by OpenAI's own policy, is the shortest-notice tier short of a preview model, and preview models (named as such in the model string) can be retired with as little as two weeks' notice. Neither Anthropic nor OpenAI has a published equivalent that guarantees the six-month runway once you're off a stable GA snapshot.

A worked example

Picture a 40-person Series A company running resume screening on a chat-latest alias, wired in eighteen months ago by an engineer who has since left, with no eval suite tracking output quality and no calendar reminder tied to either vendor's deprecation page. The retirement email lands with 94 days on the clock. Migrating means picking a replacement model, rebuilding whatever prompt tuning made the old one behave, and re-testing against real resumes before the shutdown date, all without the person who built the original integration. Ninety-four days is enough time to do that properly. It is not enough time to do it starting from a standing start with no eval harness and no owner, which is the position this company is actually in.

What this means for your team

None of this applies evenly. A team that already pins dated snapshots (not "-latest" aliases) on generally available tiers gets the full floor on both vendors, sometimes more. A team with an eval harness that can validate a replacement candidate in days, not months, barely notices a retirement notice regardless of which tier it's on. And a team intentionally running a preview model for experimentation has accepted the two-week volatility going in; that's not a gap to close, it's the deal. For everyone else:

  • Know which tier a production model ID actually sits in: dated GA snapshot, "-latest" or specialized variant, or preview. Each carries a different real notice period, from six months down to two weeks.
  • Put both vendors' deprecation pages on a recurring check. Waiting for the notification email means starting the clock at the vendor's choice, not yours.
  • Treat a "-latest" alias in production as a three-month commitment, not a permanent one, and budget migration time accordingly.
  • The actual asset that buys back runway isn't a longer notice period, it's an evaluation harness that can score a replacement model against your own cases fast enough that a vendor's floor stops being the constraint.

If your production stack is pinned to a model ID nobody has revisited since it shipped, that's usually a half-day audit, not a rebuild, and our engineering team can walk through what's actually exposed before a retirement notice forces the conversation. Our Silicon Valley team works through exactly this kind of vendor-dependency risk with engineering leaders, and we're happy to look at your specific setup. It pairs well with the broader question of whether an agent can be trusted with a longer unsupervised task and with why production AI features fail in ways a pilot never shows: a model that quietly changes underneath you is one more entry on the same list of things a pilot doesn't test for.

Sources

Frequently asked questions.

OpenAI's own deprecations page states a minimum of six months' notice for generally available models, three months for specialized variants such as chat, Codex or deep-research aliases, and as little as two weeks for preview models. Its 2026 retirements have tracked those floors closely: the GPT-5 and o3 snapshot family was announced June 11, 2026 and retires December 11, 2026, essentially six months to the day.

Anthropic's model-deprecations page states at least 60 days' notice for publicly released models. Its four most recent 2026 retirements, including Claude Opus 4.1 (announced June 5, retired August 5) and Claude Sonnet 4 and Opus 4 (announced April 14, retired June 15), landed within two days of that 60-day floor rather than comfortably above it.

Anthropic defines deprecated as still functional but no longer recommended, with a replacement and a retirement date already assigned. Retired means the model is gone: as of the retirement date, every request to it fails. OpenAI tracks the same distinction with an announcement date and a shutdown date rather than naming the stages, but the effect is identical once the date passes: requests stop working entirely.

Yes. OpenAI's own policy sets specialized variants, including chat, Codex and deep-research aliases like gpt-5.2-chat-latest, at a three-month minimum rather than the six months given to generally available snapshots; that alias was announced May 8, 2026 and retired August 10, 2026, 94 days later. Preview models, named as such in the model string, can be retired with as little as two weeks' notice under the same policy.

Track which tier a production model ID sits in, since the real notice period ranges from OpenAI's six months for a dated GA snapshot down to two weeks for a preview model, while Anthropic's 60-day floor applies uniformly rather than varying by tier. Check both vendors' deprecation pages on a recurring schedule rather than waiting for the notification email, and keep an evaluation harness ready so a replacement model can be validated in days rather than months once a retirement date is set.