Migrate OpenAI Model IDs Before the July 2026 Shutdown Wave

OpenAI's July 23, 2026 model shutdown removes preview snapshots for computer use, Codex, search, audio, realtime, TTS, and deep research. The immediate migration is mostly snapshot replacement. The larger outage risk is October 23, 2026, when gpt-3.5-turbo, gpt-4, gpt-4-turbo, o1, o3-mini, o4-mini, gpt-image-1, and related fine-tuned models shut down.[1]

The thing most teams will miss: July 23 is not the worst migration. It is the rehearsal. The July list is noisy because it includes recent-sounding model IDs such as gpt-5.1-codex and gpt-5.2-codex, but October is where hidden fallbacks break production. If your retry path still says "fall back to gpt-3.5-turbo," your primary model can be healthy while your incident handler returns 404s.[1][3]

Treat July 23 as a Snapshot Cleanup, Not a Full Platform Rewrite

The July 23 wave shuts down dated snapshots, not every capability in those families. OpenAI's deprecations page lists computer-use-preview-2025-03-11, search and TTS preview snapshots, gpt-audio-mini-2025-10-06, gpt-realtime-mini-2025-10-06, multiple gpt-5.x-codex IDs, and deep-research snapshots including o3-deep-research-2025-06-26 and o4-mini-deep-research-2025-06-26.[1]

That means the first audit is mechanical: find exact model strings, map each to the listed substitute, and run regression tests on prompts where model behavior matters. Do not rely on model age as a proxy for safety. The community reaction called the list unusually large, and one developer pointed out that gpt-4-1106-preview appears again even though it had already been listed in an earlier wave.[3] The operational lesson is simple: the deprecations page, not release recency, is the source of truth.

For most teams, July's riskiest categories are realtime/audio and Codex-like coding agents. Audio migrations need latency and voice-quality checks. Coding-agent migrations need trace-level evals: run the same repository tasks before and after the model change, then compare patch quality, tool-call count, latency, and cost. If your team does not already have this vocabulary, start with agent-evaluation before treating the replacement model as equivalent.

Replace Computer-Use Preview with GPT-5.4 Mini, Not a No-Replacement Panic

The accurate computer-use migration is: computer-use-preview-2025-03-11 goes away, but computer-use capability does not. The synthesis and OpenAI model docs point to gpt-5.4-mini as the functional replacement, and the GPT-5.4 mini model page lists computer use as supported under Responses API tools.[2]

This distinction matters because developers have been reading the deprecation table as circular: the snapshot is deprecated, the alias appears in the transition path, and the long-term substitute is easy to miss. The community thread asking whether computer use was retiring with no replacement captures that confusion.[4] The safer migration rule is: stop depending on the dated preview snapshot, move screen-control workflows to gpt-5.4-mini or gpt-5.4, keep the Responses API computer-use tool declaration, and re-run your browser or desktop task harness. Computer-use agents are sensitive to screenshot cadence, tool-call budget, coordinate precision, and recovery behavior after bad clicks.

Audit October 23 for Hidden Legacy Fallbacks and Fine-Tunes

The October 23 wave is the larger production risk because it removes the model families many teams still treat as durable defaults. OpenAI lists shutdowns for gpt-3.5-turbo-0125 via the gpt-3.5-turbo alias, gpt-4-0613 via gpt-4, gpt-4-turbo-2024-04-09, gpt-4.1-nano-2025-04-14, gpt-4o-2024-05-13, gpt-image-1, o1, o1-pro, o3-mini, o4-mini, and fine-tuned models based on gpt-3.5-turbo, gpt-4, gpt-4.1-nano, babbage-002, davinci-002, and o4-mini.[1]

The dangerous files are rarely named openai-models.ts. They are eval fixtures, LangChain defaults, notebooks, queue workers, incident fallbacks, retry middleware, old fine-tune manifests, and cost-optimized "cheap model" branches. Search for exact strings and family aliases, then partials such as gpt-3.5, gpt-4-turbo, o1, o3-mini, o4-mini, ft-, babbage, and davinci.

A concrete failure mode: your primary completion route targets gpt-4.1, which is still live, but your fallback handler on connection timeout was written two years ago to retry on gpt-3.5-turbo. On October 24, the primary succeeds, the fallback returns model_not_found, and your error-rate SLO trips because the handler now hard-errors instead of recovering. The fix is not to route around the exception — it is to promote the fallback to a current model ID before the shutdown date.

Fine-tunes need a separate plan. A fine-tuned model tied to a retiring base does not magically become a fine-tuned replacement on the new base. Budget time for re-training, eval comparison, and rollout behind a flag. The key distinction in fine-tuning is that you are migrating learned behavior, not just swapping an inference endpoint.

Put the Whole H2 2026 Deprecation Calendar on One Runbook

July and October are not isolated dates. The synthesis tracks a dense second-half calendar: August 10 for gpt-5.2-chat-latest and gpt-5.3-chat-latest, August 26 for Assistants API sunset, September 28 for older instruct and base models, December 1 for image-model consolidation, and December 11 for original gpt-5 and o3 snapshots.[1][5][6]

The Assistants API deserves special attention because it is not a model swap. OpenAI's community announcement says the beta sunsets on August 26, 2026; the migration path changes object boundaries from Assistants, Threads, and Runs toward Responses and Conversations.[5] If your app uses server-managed threads as the state container, decide what state remains server-side, what moves into your database, and how much historical context each response receives. The practical migration pattern is covered in course/openai-agents-sdk-mastery/ch01-sdk-responses-api.

Image generation has a similar trap. The research synthesis cites a December 1 consolidation from gpt-image-1-mini, gpt-image-1.5, and chatgpt-image-latest toward gpt-image-2.[6] If your product stores expected dimensions or prompt templates tuned to old outputs, schedule a visual regression pass.

Ship a Model-Lifecycle Harness Before the Next Shutdown

The durable fix is not a spreadsheet of replacement names. It is a model-lifecycle harness: model IDs in config, owner-assigned deprecation checks, eval suites for every production route, and release gates that compare behavior before and after a model change. Keep aliases and dated snapshots separate, flag every fine-tuned base model, and make OpenAI's deprecations page part of release ops. A quarterly review cadence is the minimum; monthly is better for organizations with fine-tuned models, since retraining on a new base typically needs six to eight weeks of eval time before a production rollout is safe.

A practical harness has three layers. First, a central models.config.ts where every model ID is declared and versioned — no inline string literals in service code. Second, a weekly CI script that checks each declared ID against OpenAI's deprecations feed and opens a ticket automatically when a match appears. Third, a labeled eval suite that runs representative prompts per production route and records quality scores before and after any model change, so reviewers have evidence rather than opinion when approving a swap. Model IDs treated as infrastructure config get the same review rigor as a database schema migration. That discipline eliminates the silent-breakage scenario where a fallback routes to a sunset model for weeks before an on-call engineer notices the error spike.

The practical takeaway: update July IDs, but spend most engineering time on the October audit. That is where assumptions about stable OpenAI families, cheap fallbacks, and reusable fine-tunes become production risk. For migration habits covering Responses API, tool calling, evals, and release gates, continue with OpenAI Agents SDK Mastery: Build Production-Ready Autonomous Systems.