OpenAI Model Deprecations: Build a Migration Checklist Before Shutdown

Before you begin

No previous experience required unless you choose a coding exercise. Use an account you are allowed to access; features vary by product and region.

What you will learn

• Explain the difference between deprecation and shutdown.
• Identify the models and replacement suggestions in the October notice.
• Create a low-risk model migration plan.
• Test output, tools, latency and cost before switching production traffic.

What the October 1 notice says

OpenAI's deprecations page records an October 1, 2026 notice for three API models: gpt-5.3-codex, gpt-5.4-nano and gpt-5.1. The listed shutdown date is April 1, 2027. OpenAI recommends gpt-6-sol for gpt-5.3-codex and gpt-5.1, and gpt-6-luna for gpt-5.4-nano.

Deprecation is advance notice that a model or interface is on a path to removal. Shutdown is the date after which the deprecated identifier no longer works. A deprecated model may continue responding before shutdown, but starting new work on it creates avoidable migration pressure.

Sources: OpenAI API deprecations

A replacement name is a starting point

A recommended replacement is not proof that your application behaves identically. Models can differ in output wording, instruction following, tool calls, supported parameters, speed, context limits and cost. Even a better model can break software that expects one exact format.

Inventory every place the old identifier appears: application code, environment settings, scheduled jobs, evaluation scripts, dashboards and documentation. Never paste API keys into an AI prompt or source file. Store secrets in the hosting platform's protected environment settings and rotate a key if it is exposed.

Capture a baseline before changing anything. Save representative inputs, expected properties, current latency and error rates. Remove personal or confidential data unless you have authorization and an appropriate data-handling agreement.

Test before changing all traffic

Run the old and replacement models on the same small evaluation set. Score required facts, format, tool behavior, refusal behavior, latency and estimated cost. Use cases involving code should run automated tests in an isolated environment; do not execute generated commands blindly.

If results meet the acceptance threshold, send a small share of low-risk traffic to the replacement. Watch failures and keep a documented rollback while the old model still operates. Increase traffic in stages. A rollback is a planned return to the last working configuration, not an excuse to postpone migration until the final day.

Confirm the exact shutdown date and replacement again before the change. Provider schedules can be updated. Keep the model identifier configurable so the next migration does not require editing many files.

Sources: OpenAI API deprecations

Make a one-page migration card

Write the old model, proposed replacement, owner, deadline, test set, pass threshold, rollout stages and rollback trigger. Add a communication step for anyone who depends on the output. This turns a vague upgrade into an observable change.

For a classroom example, imagine a harmless FAQ summarizer. Your pass threshold might require every date to remain correct, output to be valid JSON, median response time below a chosen limit and no more than a stated cost per 1,000 requests. The exact numbers should come from your needs, not this article.

Terms and next step

Deprecation: notice that a product is moving toward removal. Shutdown: the point when the old identifier stops working. Evaluation set: representative examples used for comparison. Canary rollout: sending a small portion of traffic to a new version first. Rollback: returning to a previous working configuration. Regression: behavior that becomes worse after a change.

Next, read 'How to Design and Improve a Machine-Learning Project' for evaluation thinking, then 'How to Verify AI Answers and Detect Mistakes.'

Try this prompt

This is a suggested exercise, not a tested guarantee of any model’s output.

Create a migration checklist from the supplied deprecation notice. Include old model, recommended replacement, shutdown date, inventory locations, evaluation criteria, staged rollout and rollback trigger. Do not include real API keys or assume the replacement is behaviorally identical.

Mini project & challenge

1. Choose one fictional application and list every place its model name might be stored.
2. Write five representative test inputs with no personal data.
3. Define pass conditions for facts, format, tools, latency and cost.
4. Design a three-stage rollout: local test, small canary and wider release.
5. Write one rollback trigger and name the person responsible for acting on it.
6. Challenge: add a calendar reminder at least one month before the shutdown date and explain why waiting is risky.

Common mistakes

• Confusing deprecation with immediate shutdown: record both notice and removal dates.
• Changing only one code file: inventory jobs, settings, tests and documentation.
• Assuming replacement output is identical: compare representative cases and formats.
• Waiting until shutdown day: migrate in stages while rollback remains possible.

Check your understanding

1. What date is listed for the October deprecation notice?
2. What shutdown date is listed for the three models?
3. What replaces gpt-5.3-codex in the notice?
4. What replaces gpt-5.4-nano?
5. What replaces gpt-5.1?
6. What is the difference between deprecation and shutdown?
7. Why is a recommended replacement only a starting point?
8. What is a canary rollout?
9. Where should API keys be stored?
10. Why keep a rollback plan?

In short

A deprecation notice gives you time to migrate safely. Inventory dependencies, confirm the recommended replacement, compare representative tasks, roll out gradually and keep a rollback path before the shutdown date.

Continue the Understand AI course →

Sources & review notes

1. OpenAI API deprecations

   October 1, 2026 notice and April 1, 2027 shutdown table checked October 3. Replacement behavior was not tested for this article.

Product information was checked on 2026-10-03. This is a selected beginner guide, not an exhaustive archive of every announcement. Recheck access, pricing and compatibility before relying on a current product detail. Supplied screenshots remain unreplicated claims.