GPT-5 and o3 models retire on 11 December 2026: what to do before the date

Trust Engine AI · Updated 12 November 2026

The models, by ID

Read from OpenAI's API deprecations page on 12 November 2026 (notice dated 11 June 2026). OpenAI adds that fine-tuned versions of these models follow the same shutdown schedule as their base models. Nearby OpenAI dates: on 30 November 2026 Agent Builder, the Evals dashboard and API, and the v1/prompts API with reusable prompt objects shut down; on 14 January 2027 the gpt-4o, gpt-4o-mini, gpt-4.1 and gpt-4.1-mini snapshots listed there go, with gpt-6-luna as the replacement. All of them are in our model retirement calendar.

Azure OpenAI has its own schedule. On 12 November 2026 Microsoft's Foundry model retirement schedule (last updated 23 September 2026) listed o3 and o3-pro as deprecated with retirement on 19 November 2026, earlier than on OpenAI's API, and gpt-5.6-sol as replacement. It listed later dates for the GPT-5 family: gpt-5 and gpt-5-nano 9 February 2027, gpt-5-pro 7 April 2027, gpt-5-mini 13 July 2027. Check that page for your deployment.

What happens if you do nothing

• The model stops answering. OpenAI's page says: “At the time of the shut down, the model or endpoint will no longer be accessible.” Unlike Google's Gemini 2.5 shutdown, OpenAI does not say that requests are routed to the replacement. Plan for failed calls, not a silent switch.
• Billing. The page says nothing about billing after the date.
• Fine-tuned models go too. Fine-tuned versions of these models shut down on the same date as their base model.
• Pinned snapshots are included. Both the aliases and the dated snapshots in the table are listed. Pinning gpt-5-2025-08-07 does not buy time.
• Defaults. If a framework or SDK picks the model for you, check which ID it sends.
• You can check programmatically. OpenAI says GET /v1/models returns an optional shutdown_date field for each model. Add it to your monitoring.

Migration checklist

1. Find every call site and model ID. Search code, config, environment variables and framework defaults for gpt-5, o3 and the dated snapshots. Note the reasoning effort setting at each call site.
2. Freeze 20–30 real tasks with expected answers. Include the cases that must never go wrong; write the answers down before any run.
3. Run the old and candidate setups on the same tasks. Your current model, the named replacement, and one alternative (for example the next tier down, or a lower reasoning setting). Same prompts, same tools, same grading.
4. Compare cost per finished task and the named tasks that change. Count retries and failures in the cost. List exactly which tasks each setup gets wrong; an average hides them.
5. Check the mechanics. Tool calls, structured output against your schema, latency, rate limits and reasoning token use on the new model.
6. Roll out behind a flag. Move a share of traffic, watch the frozen set and production, then switch fully before 11 December.
7. Keep the evidence. Store the task set, results and decision. The 14 January 2027 retirements need the same test.

Why averages mislead

In our public test of the Aider coding agent on 27 tasks with hidden tests, Claude Sonnet solved the same 26 tasks as Claude Opus at 49% lower cost per solved task. Claude Haiku was far cheaper but failed 4 more tasks, mostly edge cases. An average hides that; the named failures show whether those 4 tasks matter to you.

That test used Claude models, not OpenAI. The method is the same for any switch, including o3 or GPT-5 to the GPT-5.6 family: same tasks, known answers, cost per finished task, failures by name. Full method and raw data: Aider cost vs quality report.

Questions people ask

Questions or corrections: hello@agenttrustapi.com.