Skip to main content

Model Deprecation Tracker

Deprecation and retirement dates for major models — plan migrations before EOL.

100% client-side⎘ exportable output⌁ zero network calls
10 models

in this snapshot are deprecated or retired across Anthropic, OpenAI and Google — 15 rows match your filters. Dataset verified 2026-06-11.

Model deprecation status by vendor
ModelVendorStatusGuidance
Claude Opus 4.5AnthropicactiveCurrent flagship — no action needed. vendor page ↗
Claude Sonnet 4.5AnthropicactiveCurrent workhorse — no action needed. vendor page ↗
Claude Haiku 4.5AnthropicactiveCurrent fast tier — no action needed. vendor page ↗
Claude Opus 4.1AnthropicdeprecatedMigrate to Claude Opus 4.5 before retirement. vendor page ↗
Claude Opus 4AnthropicdeprecatedMigrate to Claude Opus 4.5 before retirement. vendor page ↗
Claude Sonnet 4AnthropicdeprecatedMigrate to Claude Sonnet 4.5; prompt changes are usually minimal. vendor page ↗
Claude Haiku 3.5AnthropicretiredRetired on the first-party API (still available via Bedrock/Vertex). Move to Haiku 4.5. vendor page ↗
GPT-5.x familyOpenAIactiveCurrent generation — no action needed. vendor page ↗
Fine-tuning platformOpenAIdeprecatedWinding down — closed to new users. Plan around prompting, RFT alternatives or other vendors. vendor page ↗
o1-preview / o1-miniOpenAIretiredReplaced by later reasoning models — migrate to the current reasoning tier. vendor page ↗
GPT-4.5 PreviewOpenAIretiredRemoved from the API — use the current flagship tier. vendor page ↗
GPT-4 (original)OpenAIretiredRetired — migrate to the current flagship tier. vendor page ↗
Gemini 3 familyGoogleactiveCurrent generation — no action needed. vendor page ↗
Gemini 1.5 Pro / FlashGoogledeprecatedNot available on new projects — migrate to the current Gemini generation. vendor page ↗
PaLM 2 (Bison)GoogleretiredFully retired — migrate to Gemini. vendor page ↗

This is a point-in-time snapshot (2026-06-11) of vendor deprecation pages, not a live feed. Always confirm dates on the linked vendor page before planning a migration.

markdown export, no lock-in
100%
generated locally
0
signup walls
0
network requests per keystroke

How it works

Every production AI integration carries a quiet dependency: the model identifier in your config will eventually stop working. This tracker holds an embedded snapshot of the deprecation and retirement status of major models from Anthropic, OpenAI and Google — verified against the vendors' own deprecation and changelog pages on the date shown in the footer — in one searchable table with filter pills by vendor and status. Each row carries concrete migration guidance and a link to the authoritative vendor page, because the snapshot tells you where to look, and the vendor page tells you the date that binds.

The status taxonomy is deliberately simple. Active models need no action. Deprecated models still serve traffic but have an announced or impending end-of-life — this is your migration window, and the time to re-run evaluations against the suggested target. Retired models fail at request time; if one of these is still in a config file somewhere, that integration is already broken or running on a vendor-side alias you do not control. The retired tier also captures partial retirements, such as models removed from a first-party API while remaining available through cloud marketplaces like Bedrock or Vertex.

Capability sunsets appear alongside model sunsets. The wind-down of OpenAI's classic fine-tuning platform is the current example: teams with fine-tuned models face exactly the same forced-migration mechanics as a model EOL, so it earns a row. A complete risk picture covers everything in your stack with a vendor-controlled shutdown switch.

Practical workflow: search for each model identifier that appears in your codebase or config, copy the filtered rows as markdown into your planning doc, and attach the vendor links as the source of truth for dates. Pair this with the context window comparison table when choosing a migration target, and with model pricing to check whether the successor changes your cost per task. The honest caveat repeats on the page itself: this is a dated snapshot, not a feed — verify before you commit a timeline.

Frequently asked questions

What is the difference between a deprecated and a retired model?

Deprecated means the vendor has announced end-of-life: the model still works, but it is closed to new use cases and has a retirement date, announced or pending. Retired means requests fail — the model identifier returns errors and your integration breaks. The window between the two states is your migration budget; vendors have historically given anywhere from two months to a year, and the trend is toward shorter windows.

How much warning do vendors give before retiring a model?

It varies by vendor and by how central the model is. Anthropic publishes a deprecation page with explicit retirement dates and has typically given several months of notice. OpenAI maintains a deprecations page with shutdown dates per model and per snapshot. Google migrates models when new Gemini generations ship, sometimes blocking new projects from old models well before full shutdown. None of this is contractual for most API customers — pin nothing to a model you cannot leave.

How should a team plan a model migration?

Treat it as a behavioral change, not a string swap. Re-run your evaluation suite against the target model first, because tone, formatting, refusal behavior and tool-calling reliability all shift between generations. Migrate one workload at a time behind a config flag so rollback is instant, compare cost per task since newer models often have different pricing, and only delete the fallback after a full business cycle on the new model.

Is this tracker a live feed of vendor deprecations?

No — it is an embedded snapshot verified on the date shown in the table footer, with a link to the official deprecation page for every row. Vendor timelines change with little notice, so use this tool to discover which of your models carry EOL risk, then confirm exact dates on the linked vendor page before committing a migration plan or telling stakeholders a date.

Why does the OpenAI fine-tuning platform appear in a model tracker?

Because platform capabilities retire just like models do, and fine-tuning is the clearest current example: the classic fine-tuning offering is winding down and closed to new users. Teams that built workflows around fine-tuned models face the same migration problem as a model EOL — they must move to prompting strategies, newer adaptation methods, or another vendor. Tracking capability sunsets alongside model sunsets keeps the risk picture complete.

FORG tracks this automatically across every agent session — live cost attribution, budgets, and alerts.

Start tracking with FORG