OpenAI's Biggest Deprecation Wave Yet: gpt-4, o1, o4-mini, and 25+ Models Shut Down by October 2026

If your team has been running on gpt-4, gpt-3.5-turbo, o1, o3-mini, or o4-mini — or any of the gpt-5.x-codex snapshots — OpenAI's April 22 deprecation notice is the most consequential model lifecycle event of 2026. Two hard shutdown dates are now on the calendar: July 23 and October 23, 2026. After those dates, API calls to the affected model IDs return errors, not responses.

What happened

On April 22, 2026, OpenAI published a batch deprecation covering more than 25 model IDs. A follow-up on May 8 added gpt-5.2-chat-latest and gpt-5.3-chat-latest (shutdown August 10). The April 22 batch splits into two waves:

Wave 1 — July 23, 2026:

Wave 2 — October 23, 2026:

The Assistants API also shuts down August 26, 2026, with migration to the Responses API.

Why it matters for AI engineering teams

This is not a routine model refresh. Several of these IDs have been production workhorses for two to three years:

gpt-3.5-turbo is still the default fallback in a large share of routing configs, cost-optimized pipelines, and fine-tuning workflows. Its October 23 shutdown means any system that falls back to it under load or on error will start returning 404s.

o1 and o3-mini are widely used for reasoning-heavy tasks. Both shut down October 23. The replacement is gpt-5.5, which has a different pricing tier and latency profile — not a drop-in swap for cost-sensitive workloads.

o4-mini shuts down October 23 as well. Teams that built agent loops around o4-mini's cost/reasoning tradeoff need to re-evaluate against gpt-5.4-mini.

gpt-image-1 shuts down October 23. If you're routing image generation requests to it, the replacement is gpt-image-2, which has different pricing and output characteristics.

Fine-tuned models are tied to their base model's deprecation date. If you have fine-tuned ft-gpt-3.5-turbo or ft-gpt-4 models in production, inference on them stops when the base model shuts down — October 23. There is no automatic migration of fine-tuned weights.

The July 23 wave is closer and catches teams off guard because it includes the gpt-5.x-codex family — models that were only released in 2025 and 2026. The assumption that "newer models have longer lifetimes" does not hold here.

The router/operator angle

Routing table audit is the first action. Every model ID in your routing config, fallback chain, and provider-specific override needs to be checked against the deprecation list. The risk is not just the primary model — it's the fallback. A system that routes to gpt-5.5 as primary but falls back to gpt-4 on rate-limit errors will silently break on October 23.

Fallback chain reconstruction. The replacement mappings are not always cost-equivalent:

• gpt-3.5-turbo → gpt-5.4-mini: pricing and latency are different; test before swapping
• o1 → gpt-5.5: significant cost increase; consider gpt-5.4-mini for tasks that don't need deep reasoning
• o4-mini → gpt-5.4-mini: reasonable swap, but verify tool-call and structured-output behavior

Fine-tuning customers face a separate migration path. Fine-tuned model inference stops with the base model. OpenAI's self-serve fine-tuning platform is also being restricted: as of May 7, new fine-tuning jobs are blocked for organizations that haven't previously run fine-tuning. By July 2, organizations that haven't run inference on a fine-tuned model in 60 days lose access to create new jobs. Plan fine-tuning migrations before those gates close.

Model ID pinning vs. alias routing. Teams using pinned snapshot IDs (e.g., gpt-4-0613) get a hard error on shutdown. Teams using aliases (e.g., gpt-4) also get a hard error — the alias is being removed, not redirected. Neither approach provides automatic continuity.

Observability gap. If your logging only captures model name and not the full model ID returned in the response, you may not know which snapshot you're actually hitting. Now is the time to add model ID logging before the shutdowns create ambiguity in incident postmortems.

What to watch on the July 23 deadline: The gpt-5.x-codex family shutdown is the one most likely to catch teams by surprise. These models were used in Codex-based coding agent workflows. If you're running any agent that calls gpt-5-codex or gpt-5.1-codex as a backend, that workflow breaks on July 23 — less than two months away.

What TheRouter users should watch or try

If you're routing through TheRouter, the model ID mapping in your routing config is the first thing to audit. Check whether any of the deprecated IDs appear as primary models, fallback targets, or per-team overrides. The July 23 deadline is 53 days away.

For teams migrating fine-tuned models: the replacement base model affects which provider endpoints are available for routing. gpt-5.4-mini and gpt-5.5 have different rate limit tiers and regional availability than their predecessors — factor that into your fallback design.

The broader pattern here is that OpenAI is accelerating its model lifecycle cadence. The gap between release and deprecation for the gpt-5.x-codex family was under 12 months. Building routing policies that assume multi-year model stability is no longer safe. Provider-agnostic routing with model-version abstraction is increasingly the right default.