OpenAI Retires gpt-5 and o3 Snapshots on December 11: What Every Routing Team Must Audit Now

Six of OpenAI's earliest gpt-5 and o3 model snapshots will stop responding to API requests on December 11, 2026. If your routing policy references any of these date-stamped IDs, you have a hard deadline six months out — not a soft migration suggestion.

What OpenAI deprecated

On June 11, 2026, OpenAI notified developers of the following shutdowns:

This is OpenAI's second major deprecation wave of 2026. The first, in late May, covered legacy models like gpt-4, o1, and o4-mini. This wave is distinct: it retires the initial public snapshots of the gpt-5 family and both o3 variants. Teams that explicitly pinned to a date-stamped ID to lock in behavior — a reasonable practice when a model launches — now need to revalidate and re-pin or move to floating aliases.

Why this is different from the May wave

The May deprecation was about sundowning pre-gpt-5 era models that were already several generations behind. Most engineering teams had already migrated or planned to.

This wave is more operationally urgent for two reasons:

First, o3 → gpt-5.5 is not a drop-in swap. OpenAI maps both o3-2025-04-16 and o3-pro-2025-06-10 to gpt-5.5 / gpt-5.5-pro. But gpt-5.5 is a chat-completion model, while o3 was a reasoning model with a separate thinking paradigm. Teams using o3 for its structured reasoning output, its system prompt behavior, or its extended context use need to re-evaluate whether gpt-5.5 with Thinking mode produces equivalent or better results for their specific workloads.

Second, the gpt-5-pro snapshot → gpt-5.5-pro replacement implies a capability-tier change. If your routing policy directs high-stakes tasks to gpt-5-pro-2025-10-06, you need to validate that gpt-5.5-pro performs at or above par on your evals before the December cutover.

The router/operator audit checklist

Before December 11, routing teams should:

1. Grep your routing configs for date-stamped IDs. Any reference to gpt-5-2025, gpt-5-mini-2025, gpt-5-nano-2025, gpt-5-pro-2025, o3-2025, or o3-pro-2025 is a migration target.

2. Separate snapshot pins from fallback chains. If you're using date-stamped IDs as fallback targets (e.g., fall back to gpt-5-2025-08-07 when gpt-5.5 is overloaded), your fallback chain will break on December 11. Update fallback targets to recommended replacements.

3. Run provider-specific capability tests for o3 → gpt-5.5. The reasoning paradigm shift means behavioral validation is required, not just latency/cost benchmarking. Test on your highest-stakes agentic and coding workloads.

4. Check downstream caching and prompt templates. Some teams cache model-specific prompt formats or system instructions. The gpt-5.5 tokenizer and instruction-following behavior differ from o3 enough to warrant a systematic prompt review.

5. Audit API cost projections. gpt-5.5 and gpt-5.5-pro are priced differently from their o3 predecessors. If your billing model is based on historical o3 rates, reforecast using gpt-5.5 pricing before the migration completes.

6. Set migration milestones. Six months sounds comfortable, but it runs through Q3 summer slowdowns and the holiday season. Plan for: validation complete by August, staging rollout by October, production cutover by November, leaving December as buffer.

What to watch on the gpt-5 side

For teams on gpt-5-2025-08-07 or gpt-5-mini-2025-08-07, the migration to gpt-5.5 or gpt-5.4-mini is closer to a drop-in — same chat-completions API, same parameter set. The main risks are:

• Instruction following deltas. Later snapshots of the gpt-5 family respond differently to system prompts that were tuned for the August 2025 behavior. Run your existing eval suite against the target model before committing.
• Latency profile shifts. gpt-5.5 may have different p50/p99 latency characteristics. If you have latency-sensitive routing rules (e.g., route to gpt-5.4-mini when the 2-second budget is exceeded), re-calibrate thresholds.

What TheRouter users should watch

Teams routing to OpenAI through a gateway should audit every named model config, routing rule, and fallback chain that references any of the six deprecated IDs. Update target model fields to the recommended replacements and validate the changes in a staging route before promoting to production. For the o3→gpt-5.5 migration specifically, a thinking-mode comparison eval is worth running before the cutover — the behavioral contract shifts enough that relying on informal smoke tests is a risk.

For any migration that spans multiple providers or model tiers, a unified audit pass through your routing policy — listing all pinned model IDs, their deprecation status, and the planned replacement — is worth completing before August.