∞Direct answer
Direct answer
Save the work your team controls before a model retires. Keep your prompts, test cases and dated results outside its dashboard. A prompt is the text that guides the model. Test a new route on real tasks. Set a safe fallback for when the old route ends. Name who owns the switch and each open gap.
Map what could stop working
Start with one core task that uses an outside AI model. List the exact model ID and the firm that hosts it. An ID is the model's name in a request. Add the API path: the route your code calls. Note any saved prompts, files or tools it needs.
Check the notice for that model, service and host. Do not treat a whole brand as one shutdown date. As read on 5 November 2026, OpenAI lists a near-term service change. Its saved prompt objects and Evals service are due to shut down on 30 November 2026. Evals became read-only on 31 October. These dates do not mean all OpenAI models end. [1]
Anthropic says calls to retired models fail. Its notice floor is 60 days for publicly released models. This applies to customers with models in active use. Partner-run hosts can set their own dates. [2] Check the host you use, not just the model name.
- Core task, exact model ID and host.
- API path, saved state and linked tools.
- Notice link, date checked and switch owner.
Keep a usable copy of your work
Store the files you may keep in your own secure space. Save prompt text and the fields that change for each task. Keep the code version, tool rules and required answer fields too. Do not copy secrets into this pack. Remove client details that the tests do not need.
OpenAI's prompt guide says to move prompt content into code. It also moves prompt versions and tests into your own repo. A repo stores code and its change history. [4] A saved prompt ID alone is not the prompt text. Make sure the team can run without that ID.
Save test inputs, expected results and dated past outputs. If a test tool ends, rebuild its checks elsewhere. OpenAI's Evals move guide shows how to rebuild tests in Promptfoo. It does not require an Evals export feature. It warns that scoring may differ. Check the rebuilt tests before you trust them. [5]
- Prompt text, changing fields and code version.
- Test inputs, expected results and past outputs.
- Scoring rules, source files and access limits.
Test the whole task, not just the model
Choose cases from the work users need done. Include common tasks, hard cases and unsafe requests. OpenAI advises task-specific tests and human checks. AI can give different outputs for the same input. [3] One good demo is not enough proof of a safe switch.
Run old and new routes on the same allowed test set. Check facts, answer fields and tool actions. Test each language your product serves. Check what happens when a file is missing or a call fails. A new model name does not prove that the full task works.
Measure cost and time on that same task set. Count failed calls and retries too. A retry is a call made again after a failed attempt. Write the test date, model IDs and test size beside each result. Keep a person in charge of hard judgments.
- Same task inputs and clear pass rules.
- Facts, fields, tools and safe refusal checks.
- Cost, time, failures and reviewer notes.
Set a hold rule and a fallback
Set your own pass rules before the test starts. There is no pass score here for every startup. A leak of private data should not vanish in an average. Hold the switch if a core safety check fails. Fix it and test again before expanding use.
Try a small release after the checks pass. Set who can stop it and what would cause a stop. A return to the old model only works while it is live. Once it retires, that is not a fallback. [2]
Plan another tested route or a limited manual mode. A manual mode puts the task with a person. Tell users which parts are paused. Do not send their data to a new host without approval. Check data terms and local law before changing where private files go.
- Hold rule for core safety or task failures.
- Small release, stop trigger and named owner.
- Tested fallback after the old route ends.
Try a support-answer switch record
This is an illustrative example, not a real startup result. A startup uses AI to draft answers from its support guide. Its founder plans a move before the current model retires. The team keeps the guide, prompt and test set in its own secure space.
The tests ask for a correct answer and a source link. An unclear case must go to a person. The team checks that the new route follows those rules. It checks that a missing guide leads to a safe pause. No cost, speed or pass rate is claimed here.
If the new route invents a refund rule, hold the switch. Keep staff drafting answers until the issue is fixed. Record the failed case, owner and next test date. A clean record states what failed, not just that a test ran.
- Task: draft an answer from the approved guide.
- Proof: dated cases, source links and review notes.
- Gap: invented rule; hold and retest.
Show investors proof and open gaps
Due diligence means checks before you invest. Give investors a short, source-linked exit record. State which task depends on which host. Show the saved files, test date, result and fallback. Mark untested claims as open, not complete.
At pre-seed, start with the task your product most needs. At seed, link tests to a shipped code version. At Series A, name who repeats checks as tools change. These are steps to plan, not fixed rules for each funding stage.
Keep the notice link and next check date in the record. Ask the tech lead to review the switch plan. Ask a qualified local lawyer about hard data or contract questions. A tested exit reduces one risk. It does not remove every risk from outside firms or tools.
- Claim, dated proof and open gap.
- Named owner, next check and shutdown trigger.
- Fallback scope and limits users will see.
FAQ
Frequently asked questions
Does deprecated mean the model is already off?
Not always. Anthropic uses deprecated for a model that still works. Retired models are no longer available, and requests fail. [2] Check the exact model, host and date.
Can we keep just the old prompt ID?
No. Keep the prompt text and changing fields you may store. OpenAI's move guide puts prompts into product code. [4] Do not assume a saved ID will work after its service ends.
Can we copy the hosted model itself?
Do not assume that. This plan saves your allowed work and evidence. It does not grant rights to a vendor's model files. Check the contract and available tools with the provider.
What should an investor see?
Show the core task, host, dated tests and open gaps. Add the switch owner and a tested fallback. A vendor's suggested model is a starting point, not proof your task works.
REF
Sources
- 1
- 2
- 3
- 4
- 5
OpenAI — Konstantine Kahadze
Moving from OpenAI Evals to Promptfoo — 3 June 2026 ↗