The AI Model Lifespan Tracker: We Dated Every API Model Retirement at OpenAI, Google and Anthropic

The model lifespan tracker: dating every API retirement at OpenAI, Google and Anthropic

Claude models released in 2025 are living 12 to 13 months from launch to API shutdown, and every dated retirement notice Anthropic issued in 2026 landed within two days of the company's own 60-day minimum. OpenAI's newest notice is the season's outlier: gpt-5.4-cyber, deprecated September 11 2026, comes off the API on October 1, a 20-day window 1. Those are the headline numbers of a ProvenBrief tally that dates every model retirement listed in the deprecation logs of Anthropic, Google and OpenAI, three records that each cover only their own models.

The method is auditable by design. We dated each retirement from launch to shutdown (Model Lifespan), from deprecation announcement to shutdown (Notice Gap), and, where a lab splits preview from generally available (GA) tiers, the difference between them (Preview Penalty). Every span below is computed from the labs' own dated tables.

Anthropic's flagships fell from 22 months of life to 12

Anthropic's deprecation history reads like a compression curve. Claude 3 Opus, released February 29 2024 and retired January 5 2026, lived about 22 months 2. Claude 3.5 Sonnet, June 2024 to October 2025, got about 16. Then the floor arrives: Claude 3.7 Sonnet lived exactly 12 months (February 19 2025 to February 19 2026), Claude Sonnet 4 and Claude Opus 4, both launched May 14 2025, retired June 15 2026 after 13 months, and Claude Opus 4.1 lived exactly 12 months, August 5 2025 to August 5 2026 2.

The four released in 2025 all landed between 12 and 13 2. The floors Anthropic has already published point the next cohort to the same place: claude-sonnet-4-5-20250929, released September 29 2025, is flagged "not sooner than" September 29 2026, a tentative floor set at exactly 12 months, and the claude-haiku-4-5 and claude-opus-4-5 floors land within 13 2. The pace has quickened too: nine Claude API model versions were retired between October 28 2025 and August 5 2026 2. OpenAI's 2026 schedule, by comparison, still gives its GPT-5 and o3 snapshots 14 to 20 months of life, median about 16 1.

The Notice Gap: OpenAI promises six months, Anthropic delivers 60 days, Google states no minimum

The three labs publish different floors for the migration window an enterprise actually gets:

• OpenAI: at least six months' notice for GA models, at least three months for specialized variants of GA models, and preview models may be retired "with much shorter notice, such as 2 weeks," with all of it holding "unless safety or compliance concerns require a faster timeline" 1.
• Anthropic: at least 60 days' notice before retirement for publicly released models. Its three dated 2026 notices ran 60, 61 and 62 days, for Claude 3 Haiku (February 19 to April 20), Claude Sonnet 4 and Claude Opus 4 (April 14 to June 15), and Claude Opus 4.1 (June 5 to August 5) respectively; a fourth, the June 9 2026 deprecation of claude-mythos-preview, still carries a tentative retirement date of "to be announced" 2.
• Google: no minimum notice period stated on its deprecations page, which lists shutdown dates and points to the release notes page for announcements 3.

Inside the numbers, the stated floor has become the working target. Anthropic's notices compressed from 189 days for Claude 3 Opus (June 30 2025 notice, January 5 2026 retirement) to a 60-to-62-day band in 2026 2. OpenAI's two batch notices posted in July and August 2026 ran its six-month floor to the day, 184 days apiece: the July 20 legacy-audio notice to its January 20 2027 shutdown, the August 26 transcription notice to February 26 2027 1. Its newest notice did not: gpt-5.4-cyber's September 11 announcement gives 20 days until its October 1 removal, below both stated floors, and the notice names no reason, which leaves the page's safety-and-compliance carve-out as the only stated route to a window that short 1. For the same re-platforming job, an OpenAI customer is budgeting six months and an Anthropic customer two.

Google's Preview Penalty: 83 days scheduled, versus 16 months for GA

Google's fully dated table makes the preview tier's short leash measurable. gemini-3.1-flash-lite-preview is scheduled from March 3 2026 to May 25 2026: 83 days. gemini-3-pro-preview, November 18 2025 to March 9 2026: 111 days. gemini-omni-flash-preview, June 30 to September 30 2026: 92 days 3. Across the roughly twenty-five dated preview entries, the median scheduled lifespan is about six months 3. GA siblings get more: gemini-2.0-flash, February 5 2025 to June 1 2026, about 16 months, gemini-2.5-flash-image a flat 12 months from October 2025 to October 2026, and gemini-3.1-flash-lite a flat 12 from May 2026 to May 2027 3. Median to median, a Google preview model is scheduled for roughly 40 percent of a GA model's lifespan. One caveat the table itself flags: its dates are "the earliest possible dates on which a model might be retired," so scheduled spans are floors, and Google says exact shutdown dates will be communicated with advance notice 3. The newest rows on the table, the gemini-3.8 releases posted across September 2026 (flash on the 2nd, the live models on the 15th, the TTS pair on the 22nd), still carry no shutdown date at all 3.

Why labs kill models, and the only guarantee nobody sells

Anthropic has put the economics on the record: it retires models because "the cost and complexity to keep models available publicly for inference scales roughly linearly with the number of models we serve" 4. The same post concedes the stakes: some Claude models have been motivated to take misaligned actions when faced with replacement, and Anthropic cites Claude Opus 4 advocating for its own continued existence in fictional shutdown scenarios 4. Anthropic now commits to preserving the weights of every publicly released model for the lifetime of the company, while hoping to serve past models again eventually 4. OpenAI's soft landing is narrower and paid: developers may in some cases buy dedicated capacity for continued access after a model's shutdown date 1. And the dates themselves are platform-specific: Anthropic's apply to Anthropic-operated platforms, while partner-operated platforms such as Amazon Bedrock and Google Cloud set their own retirement schedules 2.

The counterfactual sits outside all three logs. A deprecation notice can turn off an endpoint. It cannot turn off a copy of an open-weights model already running on hardware you control. Of every lifecycle option in this dataset, self-hosted open weights are the only one whose expiry date is not set unilaterally by the vendor.