OpenAI Retires Codex, Deep-Research, and Computer-Use Models on July 23: The 21-Day Routing Deadline Every Team Must Clear

Three weeks from today — July 23, 2026 — OpenAI will shut down a set of model aliases that power many active coding-agent and deep-research pipelines. This is not the October 23 retirement wave that took GPT-4 and o1. It is an earlier, separate deadline covering the Codex, deep-research, and computer-use model families that many teams route to by name. If your gateway or SDK configuration still references any of them on July 24, every request will fail with a model-not-found error.

What shuts down on July 23

The following aliases and snapshots reach their end-of-life date on July 23, 2026. All return errors after that date; no automatic redirect to a replacement occurs.

Two categories deserve special attention for routing teams.

Codex variants: gpt-5-codex, gpt-5.1-codex, gpt-5.1-codex-max, gpt-5.1-codex-mini, and gpt-5.2-codex are all removed. These model IDs were commonly routed to by agent frameworks that hardcode the "Codex" naming convention. If your routing policy specifies any of these by name — in a gateway model allowlist, in a Cursor or Claude Code API override, in an OpenAI-compatible proxy config — the alias stops resolving.

Deep-research variants: o3-deep-research and o4-mini-deep-research both go to gpt-5.5-pro. If your agent pipeline triggers a deep-research model via a model alias, you are pointing at a soon-to-be-invalid endpoint. The upgrade target (gpt-5.5-pro) has meaningfully different pricing and token-budget behavior, which means this is not a drop-in alias swap — it requires a cost and context-window audit.

What else changed today (July 2)

Today, OpenAI enforced the next step of its self-serve fine-tuning phase-out. As of July 2, 2026, organizations that have not run inference on a fine-tuned model in the past 60 days can no longer create new fine-tuning jobs. If your team uses custom fine-tuned checkpoints in a routing policy — routing to ft:gpt-* style model IDs — confirm your inference logs show active usage in the past 60 days, or your ability to create replacement jobs for the October 23 base-model retirements disappears before you need it.

The timeline for fine-tuning:

• Today, July 2: No new fine-tuning jobs for orgs without recent inference usage.
• January 6, 2027: Active customers lose the ability to create new fine-tuning jobs entirely.
• Model-dependent: Inference on existing fine-tuned models continues until the underlying base model is deprecated (e.g., ft-gpt-3.5-turbo inference ends October 23, 2026 when gpt-3.5-turbo retires).

Why this matters for AI engineering teams

The July 23 wave is operationally distinct from the more publicized October 23 wave in two ways.

First, the July 23 models are heavily concentrated in coding-agent and autonomous-research pipelines — the very workloads most likely to have hardcoded model references. Agent frameworks built on Codex variants often use the model name as a capability signal, not just a routing target. Changing the model name means re-evaluating whether output quality and behavior remain stable, not just updating a string.

Second, the gpt-5.4-mini replacement for computer-use-preview and search-preview models represents a real capability change. computer-use-preview was a specialized task model; gpt-5.4-mini is a general-purpose mini model. If you route computer-use requests to the retired model, you cannot simply redirect to gpt-5.4-mini without validating that the downstream tasks still work. A browser-automation or desktop-control pipeline that worked on computer-use-preview needs explicit regression testing against gpt-5.4-mini.

The router and operator angle

For teams using an AI gateway or routing proxy, the July 23 deadline creates three concrete action items.

1. Audit every model name in your routing config today. Pull every model ID referenced in your gateway allowlist, fallback chain, and per-route overrides. Flag any alias from the table above. Do not rely on the assumption that a gateway auto-redirects deprecated model names — most do not.

2. Update fallback chains before the deadline, not after. A common failure pattern: a primary model is updated to gpt-5.5, but the fallback still points to gpt-5.1-codex. On July 24, when the primary hits a rate limit and the gateway falls back, the request fails. This is especially dangerous in Cursor, Claude Code, or OpenCode configurations where the fallback model is set separately from the primary.

3. Re-price your deep-research lane. If you route to o3-deep-research or o4-mini-deep-research, the replacement is gpt-5.5-pro. Confirm your per-request budget and token accounting still hold. gpt-5.5-pro has different input and output pricing than either deep-research alias, and the context window and thinking-budget semantics differ.

What TheRouter users should watch

Review your routing policy's model references against the July 23 table above. For gateway configs managed through TheRouter, update the model name in each route that references a retiring alias and confirm the fallback chain uses a valid replacement. For coding-agent setups using Cursor or Claude Code pointed at an OpenAI-compatible gateway, check the apiModel or model override field in your project settings.

The October 23 wave — which retires gpt-3.5-turbo, gpt-4, gpt-4-turbo, gpt-4o-2024-05-13, o1, o4-mini, and dozens more — is the larger and more disruptive event, but July 23 arrives first and with less public attention. Three weeks is enough time to audit and update routing config, but not enough time to absorb a surprise outage on a deadline you missed.

Primary source: OpenAI Deprecations