AWS Bedrock & Azure Foundry Retirement Calendar

Every announced Bedrock EOL and Azure Foundry retirement, with the action per row. Last verified: 2026-07-17. Updated weekly. Sources: Bedrock lifecycle · Azure Foundry schedule.

The two big cloud LLM platforms retire on their own schedule, independent of the upstream model vendor - a model can live on the Claude API long after Bedrock or Azure drops it (or vice versa). Two platform-specific mechanics matter here: Bedrock's "extended access" stage (the model keeps serving but the price can rise at any time, with a later hard EOL), and Azure's habit of retiring preview snapshots on tight cycles. Both are hosted here together because most teams run at least one of them as infrastructure, not as a model choice.

AWS Bedrock - model end-of-life

DateWhat retires (EOL)Do this before the dateSrc
Jul 30Claude 3 Sonnet; Claude 3.5 Sonnet (20240620, select regions + GovCloud); 3.5 Sonnet v2 (AP regions); 3.7 Sonnet (GovCloud)Repoint to Sonnet 4.6 or Sonnet 5. Check per-region: the same model EOLs on different dates by region, so a multi-region deployment can partly break while other regions still serve.
Jul 30Agents Classic closed to new customers; model catalog frozen (existing agents keep running for now)If you are still building on Agents Classic, stop - new work should target AgentCore. Plan the port of existing Classic agents before the catalog freeze forces a stale model on you.
Aug 19Cohere Command R, Command R+Migrate to Cohere Command A. Re-test RAG/tool-use prompts - Command A changes tokenization and tool-call formatting.
Aug 26Price risk: Jamba 1.5 Large/Mini enter extended access (price may rise; hard EOL Nov 26)Extended access means the meter can change with little notice. Either accept the price risk with eyes open, or migrate to a newer Jamba now. Set a reminder for the Nov 26 hard cutoff.
Sep 10Claude 3 Haiku (incl. GovCloud)Move cheap/high-volume paths to Haiku 4.5. Re-check latency and cost at your volume before assuming parity.
Sep 14Amazon Nova Premier, Nova SonicRepoint to a newer Nova. Nova Sonic (speech) has no exact drop-in - test the replacement's audio quality before cutover.
Sep 30Amazon Nova Canvas, Nova Reel (v1:0, v1:1) - image/videoNo listed first-party successor. Evaluate an alternative image/video model now and abstract the call site so the swap is contained.
Oct 8Price risk: Claude Opus 4.1 enters extended access (hard EOL Jan 8, 2027)Migrate to Opus 4.8 to escape the price-risk window, or budget for a possible rate rise. Jan 8 is the hard stop.
Oct 14Claude Sonnet 4 (all regions; in extended access since Jul 14)Repoint to Sonnet 4.6 or 5. If you deferred this at the Jul 14 extended-access start, this is the hard EOL - do not defer again.
Nov 26Jamba 1.5 Large/Mini (hard EOL)Final cutoff after the Aug 26 extended-access start. Must be migrated by now.
Nov 30TwelveLabs Marengo Embed v2.7If you built a vector index with this embedding model, you must re-embed your corpus with the successor - embeddings are not cross-compatible. Plan the re-index compute and downtime.
Jan 8, 2027Claude Opus 4.1 (hard EOL)Final cutoff after the Oct 8 extended-access start. Move to Opus 4.8.

Azure AI Foundry - model retirements

DateWhat retiresDo this before the dateSrc
Jul 23gpt-realtime-mini (2025-10-06)Move to version 2025-12-15. Realtime deployments pin a dated version - bump the deployment, do not just change the model name.
Jul 31Stable-Diffusion-3.5-Large, Stable-Image-Core, Stable-Image-UltraNo listed successor on Foundry - evaluate another image model or a different platform.
Aug 5claude-opus-4-1 (preview)Move to claude-opus-4-8.
Aug 5 / 28gpt-chat-latest preview snapshots (05-05 to Aug 5; 05-28 and 06-24 to Aug 28)Move to a newer snapshot. Stop pinning preview -latest snapshots in production.
Aug 13DeepSeek-R1 (Legacy)Repoint to DeepSeek-V4-Pro.
Aug 15MAI-Image-2e, MAI-Image-2 (preview)Move to MAI-Image-2.5 or MAI-Image-2.5-Flash.
Aug 17mistral-document-ai-2505Move to version 2512 or mistral-ocr-4-0.
Aug 30model-router previews (2025-05-19, 2025-08-07)Switch to model-router GA.
Aug 31TimeGEN-1 (Nixtla); tsuzumi-7b (NTT)Move to TimeGPT; tsuzumi2 respectively.
Sep 15sora-2 (2025-12-08 preview)No direct successor listed for this snapshot - see the Sep 24 OpenAI Sora shutdown; plan video off first-party.
Sep 16 / 18o1 (2024-12-17); o1-pro (2025-03-19)Move to o3 or GPT-5.x. Re-test reasoning-heavy prompts across the jump.
Sep 21gpt-audio-mini (2025-10-06)Move to version 2025-12-15.
Oct 1gpt-4o (all 2024 versions); gpt-4o-mini; o3-miniBig one for legacy apps. Move to gpt-5.1 / gpt-4.1-mini / o4-mini. Re-test - 4o to 5.x changes behavior, cost and token accounting.
Oct 13Cohere-command-a-plus-05-2026 (preview)No listed successor - check for a GA Command A on Foundry.
Oct 14gpt-4.1, gpt-4.1-mini, gpt-4.1-nanoMove to GPT-5.x. If you migrated off 4o onto 4.1 as a bridge, note 4.1 retires soon after 4o - go straight to 5.x.
Oct 15gpt-4o-mini-transcribe/tts, gpt-4o-transcribe(-diarize)Move audio/transcription to the 2025-12-15 versions.
Oct 16o3, o4-mini (2025-04-16)Move to GPT-5.x.
Oct 19claude-haiku-4-5, claude-opus-4-5, claude-sonnet-4-5 (previews)Move to the newer Claude snapshots on Foundry (the non-preview GA deployments).
Oct 23gpt-image-1 (preview)Move to gpt-image-2.
Nov 15codex-mini (2025-05-16)No listed successor - check current Codex/coding model options on Foundry.
Dec 5-26Batch: gpt-5.1-codex-max (Dec 5); o3-pro (Dec 10); gpt-5.2 (Dec 12); tts/tts-hd/whisper + several -2025-12-15 versions (Dec 15); gpt-image-1.5 (Dec 16); o3-deep-research (Dec 26)A dense December. Audit deployments in November and bump each to the newer version - do not leave a cluster of these for the holidays.

Azure Foundry - 2027 (announced so far)

DateWhat retiresDo this before the dateSrc
Feb 6gpt-5, gpt-5-mini, gpt-5-nano (2025-08-07)First-gen GPT-5 snapshots retire on Foundry too - move to a current GPT-5.x version.
Feb 23 - Mar 18gpt-realtime-1.5 & gpt-audio-1.5; gpt-5.3-codex; gpt-audio/gpt-realtime; gpt-5.4(-pro); gpt-5-codex; gpt-5.4-mini/nanoRolling realtime/audio/codex refresh - bump each deployment to its newer version.
Apr 7 - 23gpt-5-pro, gpt-image-1-mini; text-embedding-3-large/small, ada-002; gpt-image-2; gpt-5.5Embeddings alert: retiring text-embedding-3 and ada-002 means you must re-embed and re-index your vector store on the successor before this date. Plan the re-index compute early - this is the highest-effort item in the 2027 set.
May 6 / 15gpt-realtime-2; gpt-5.1 + gpt-5.1-codex(-mini)Bump to newer versions.
Jun 25 / Jul 9gpt-realtime-2.1(+mini); gpt-5.6-sol/terra/lunaBump to newer versions.
Oct 1-16Fine-tuned deployments: gpt-4o(-mini); gpt-4.1 family; o4-miniFine-tuned deployments on these bases retire - retrain your tunes on a supported base well ahead of the window.

Bedrock and Azure Foundry version independently of each other and of the upstream model vendor. Extended access (Bedrock) means the model still serves but the price can change - treat it as a migrate-now signal, not a reprieve. This calendar is maintenance guidance, not legal advice; confirm at the source before acting.

A date on this page just got closer

The free weekly digest flags every new Bedrock and Azure Foundry deadline the week it lands - retirements, extended-access price flags and breaking changes, each with a one-line action.