AI model deprecations: every shutdown date that matters

Shutdowns still ahead in 2026

Each model below has an announced end date from its provider. The migration column is the provider's official pick. Where benchr's pick differs (usually on price), the linked page explains why.

The 2026 retirement timeline

1. Mar 9Gemini 3 Pro Preview

   Shut down on the Gemini API after under four months. Vertex AI listed it discontinued by March 26.

2. Apr 20Claude Haiku 3

   Retired. The last Claude 3-generation model to go.

3. May 12DALL-E 2 and 3

   Removed from the API in favor of GPT Image 2.

4. Jun 1Gemini 2.0 Flash family

   Shut down. Google's recommended path is Gemini 3.5 Flash.

5. Jun 15Claude Sonnet 4 + Opus 4

   Both retire on the Claude API. The Claude 4.x line replaces them at the same or lower price.

6. Jun 25Gemini 3 Pro Image Preview

   The preview id shuts down; the stable gemini-3-pro-image replaces it.

7. Jul 23OpenAI's pre-GPT-5.4 cleanup

   Five Codex IDs, both deep-research models, computer-use-preview, and two chat snapshots retire together.

8. Aug 5Claude Opus 4.1

   Retires exactly one year after its August 5, 2025 release.

9. Oct 16Gemini 2.5 Pro + Flash

   Google's biggest cleanup so far. Replacement pricing runs higher. Check the math first.

10. Oct 23The GPT-4 era ends

    Ten models go at once: GPT-4o, GPT-4, GPT-4 Turbo, GPT-3.5 Turbo, GPT-4.1 nano, GPT Image 1, o1, o1-pro, o3-mini, o4-mini.

How each provider retires models

Anthropic labels models active, legacy, deprecated, then retired, and commits to at least 60 days' notice for publicly released models. Each active model carries a published "not sooner than" date: Opus 4.8, for instance, won't retire before May 28, 2027. Dates apply to the Claude API, AWS, and Microsoft Foundry; Bedrock and Vertex run their own schedules. Anthropic has also committed to preserving model weights and says it hopes to make past models available again someday.

OpenAI gives at least six months for generally available models, three for specialized variants, and as little as two weeks for previews. The April 22, 2026 announcement that ends the GPT-4 era was the largest single deprecation notice any provider has issued — and it didn't stop at models. A follow-up on June 3 scheduled three platform products for November 30: the Reusable Prompts API, the Evals dashboard and API, and Agent Builder (whose official path forward is the Agents SDK).

Google publishes earliest-possible shutdown dates rather than firm ones, then confirms exact dates closer in. Preview models can go fast. Gemini 3 Pro Preview lasted 16 weeks.

Retired in the last year

Don't get burned: the 15-minute audit

You'd be surprised how many production systems still pin a model ID from 2024. Three checks cover you:

1. Grep your codebase for every model string in the tables above, including date-suffixed snapshots like gpt-4o-2024-05-13 and claude-sonnet-4-20250514. Aliases hide in config files and environment variables, not just code.

Already seeing failures? A retired ID surfaces as model_not_found on OpenAI and not_found_error on Anthropic — the error database has the per-provider diagnosis and code fixes.

2. Pull your usage export. Anthropic's console has a per-key, per-model CSV export. OpenAI's usage dashboard breaks down by model. Five minutes tells you which deprecated IDs still get real traffic.

3. Price the replacement before you swap. Same-provider upgrades aren't always neutral. Gemini 2.5 Flash users face a 5× input price jump, while Claude Opus 4 users save two-thirds. Run your token volumes through the cost calculator before picking the target, and check the current rankings in case a different provider now fits better.