Capability
Frontier Model API Lifespan by Release Date
Elapsed calendar days from first documented public API availability, plotted at each endpoint’s release date. Completed shutdowns are observed endpoints; models still callable at the 2026-08-27 cutoff are shown as right-censored ages rather than treated as finished lifespans.
Frontier Model API Lifespan by Release Date
- OpenAI
- Anthropic
Among each provider’s latest plotted release, OpenAI has the largest completed or right-censored age at 385 days.
Methodology
Unit of observation: a named, provider-hosted generative model endpoint from OpenAI, Anthropic, or Google's Gemini API that official material presented as a leading, state-of-the-art, most-capable, frontier, or direct major successor model. Economy-only, embedding, reranking, moderation, image-only, audio-only, fine-tuned, partner-hosted, and legacy pre-2023 launches are excluded. Public preview and restricted API tiers qualify when the provider explicitly said developers could call the model; a research announcement without API access does not.
Start date: the first exact date on which an official launch post or API release note says the named endpoint was available to external API developers. The date embedded in a model identifier is not treated as availability unless an official page independently establishes it. This matters for Claude 3.7 Sonnet and Claude 4, whose endpoint identifiers predate their public launch announcements. GPT-4's initial waitlist release qualifies because OpenAI said it would begin inviting developers that day; it is not mislabeled as general availability.
End date: for completed endpoints, the provider’s completed API shutdown or retirement date is used and the elapsed value is end date minus release date in UTC calendar days. For endpoints still callable at the cutoff, the value is age through 2026-08-27 and the point is marked right-censored; no eventual lifespan is assigned.
Censoring is represented rather than discarded: the active GPT-5 endpoint is plotted at its first documented API availability date with a censored point marker and age-to-cutoff. Completed and censored points must not be compared as if they had equal observation status.
Same-release-day cohorts keep the longest-lived qualifying endpoint per provider and release date, preventing duplicate timestamps from turning one release cohort into an arbitrary vertical stack. The selected endpoint is named in the point note.
Conflict handling is fail-closed. OpenAI's current deprecation page contains conflicting completed and future dates for gpt-4-1106-preview, so that endpoint is excluded. Google's current lifecycle table and release notes do not agree on the first availability of gemini-2.5-pro-preview-03-25, so it is excluded from the 2025-12-02 cohort. Claude 3.5 Sonnet uses the launch page's displayed 2024-06-21 availability date while retaining the exact endpoint identifier claude-3-5-sonnet-20240620 in the note.
Google states that dates in its lifecycle table are the earliest dates on which a model might be retired and that exact timing is communicated separately. This chart uses a Google date only after it is in the past and the current official page categorizes the model among already-shutdown models. The two Google observations should still be read as official lifecycle dates rather than independently measured outage timestamps.
Every CSV row links to a first-party availability source and names the lifecycle source when a second URL is required. Release dates are never inferred from model identifiers alone; completed endpoints use verified shutdown dates, while active endpoints are explicitly right-censored.
Frequently asked questions
What does model lifespan mean in this chart?
The x-axis is the first documented public API availability date. A completed point measures days until provider shutdown or retirement; a right-censored point measures age through the cutoff because the endpoint remained available.
Why does frontier-model API lifespan matter to the AI industry?
Endpoint longevity affects migration planning, evaluation reproducibility, and the maintenance burden for products built on hosted models. These observations document completed historical lifespans; they do not predict how long a current model will remain available.
How are still-live models shown?
A still-callable endpoint is plotted at its first documented API availability date with its age through the observation cutoff and a right-censored marker. Its eventual lifespan is unknown, so it is not treated as a completed shutdown.
Why is there only one point when several models shut down together?
Batch retirements can create multiple same-provider observations at one timestamp and overweight administrative cleanups. The chart keeps the longest-lived qualifying endpoint in each provider-date cohort; ties use the higher-capability endpoint. The point note identifies the selected model.
Does a preview model count as publicly available?
Yes, if an official source says external developers could call it through the provider API on an exact date. Restricted usage tiers and waitlists are retained and labeled. Internal tests, research demos, and announcements without API access are excluded.
Can provider lifespans be compared as policy commitments?
No. The sample mixes previews and production-oriented snapshots, providers use different lifecycle terminology, and the cohort rule selects completed events only. It documents historical endpoint availability, not a guaranteed support window.
Why are Mistral, Cohere, and other providers not plotted?
Their official lifecycle materials were reviewed, but no additional endpoint passed the same frontier-scope, exact public API availability, completed endpoint shutdown, and conflict-free date requirements during this review. They can be added later only when both dates are supported by first-party evidence.