Switching back to the old model sounds like a safety net. It stops being one when the supplier retires that model.
The old model is not a permanent safety net
A rollback plan sounds reassuring until the thing you intend to roll back to stops existing. A hosted AI model is not a spare laptop in your cupboard. Your supplier controls whether the old endpoint will still answer requests. Saving yesterday's configuration does not preserve yesterday's service. If your recovery plan depends on a retired model, it can fail even when nobody has changed your application.
That distinction matters now. Anthropic's model lifecycle documentation, checked on 16 November 2026, lists Claude Sonnet 4.5's native API identifier for retirement on 30 November 2026. It says requests to retired models fail. It also distinguishes Anthropic-operated services from partner-operated platforms, which set their own schedules. A model name without its hosting route is therefore not enough to establish your recovery window.
Consider an illustrative accounts team that moves its invoice classification to a replacement model on 25 November. Its change record says that any serious mistake will trigger a return to the previous model. That may be possible during the first few days, subject to the old route remaining available. It is not a dependable December recovery plan if the old route retires on 30 November. A fault discovered during month-end processing could arrive after the escape hatch has closed.
The business question is not simply whether the replacement works today. It is what happens if it fails next week. Put an expiry date against every recovery route that depends on a supplier's continued support. That date belongs in the business acceptance decision, not only in a developer's calendar. Treat it as the last point at which one particular option may remain usable, never a promise that the option cannot fail earlier.
Saving yesterday's configuration does not preserve yesterday's service.
Separate rollback from continuity
Rollback and continuity solve different problems. Rollback restores a previous technical state. Continuity keeps an essential business service operating when the preferred technical state is unavailable. You may be able to restore the old prompt, model identifier and application version while still being unable to complete a single request. The configuration has returned; the supplier's retired service has not.
Make the distinction visible in a short recovery record. Name the workflow, its business owner, the current model and hosting route, the previous route, the supplier's stated retirement date, and the last date you will rely on that route. Alongside those fields, describe an independent way to keep operating. Depending on the work, that might mean a tested alternative model, manual processing, a deterministic rules engine, or temporarily accepting requests without producing an immediate answer.
What this means in practice is deciding which service you are protecting. For an internal briefing generator, delaying the briefing may be acceptable. For an order-processing workflow, customers may still need an acknowledgement and an accurate record of their request. Pausing the AI step must not discard the request or imply that an order has been accepted. The continuity process needs its own permissions, records and responsible person.
Do not disguise a second untested AI service as resilience. A different supplier can interpret prompts differently, use different data-handling arrangements and return information in a different format. If that alternative has never completed the real workflow safely, it is a candidate, not a recovery route. The same applies to a manual process that depends on one unavailable employee. Ask someone other than the person who wrote the plan to explain how work would continue. If the explanation requires guesswork, the plan still needs work before a retirement deadline makes the weakness urgent.
Rollback restores a technical state. Continuity keeps the business service operating.
Test the recovery route beyond the model call
A successful model response is not proof of a successful recovery. Your business process also depends on prompts, stored instructions, data retrieval, output validation and the systems that receive the result. A fallback can return fluent text while breaking a required field or skipping an approval step. The useful test is whether a complete representative job reaches the correct business outcome.
There is a concrete dependency warning in OpenAI's current deprecations page. As checked on 16 November 2026, it lists 30 November 2026 shutdowns for reusable prompt objects and the prompts endpoint, the Evals dashboard and related evaluation services, and Agent Builder. These are platform dependencies rather than one model identifier. A rollback that relies on an expiring prompt store or evaluation service can be incomplete even if its chosen model remains available.
Build a small recovery exercise around representative examples. Include a straightforward request, missing information, conflicting evidence and an input that should be refused or referred. Use synthetic or appropriately authorised test data. Check the complete chain: the instructions load, the right information is retrieved, the output passes validation, a person can approve it where required, and the final record says what actually happened. Keep tests isolated from real orders, emails and payments unless a controlled live exercise has specifically authorised those effects.
Record how long the exercise takes and who has to intervene. An overnight recovery that requires four hours of staff attention is not an instant fallback. Preserve the approved prompts, test cases and acceptance results in storage you control, with suitable access restrictions and retention rules. Existing version control and testing tools may already cover the need. Buying or building another platform is not the first requirement; knowing which dependencies survive the change is.
A fallback that answers a prompt but cannot finish the workflow is not a working recovery route.
Put an owner and a stopping rule on the transition
Technical teams can confirm that a replacement endpoint responds. The workflow owner must decide whether the resulting business behaviour is acceptable. Keep those responsibilities connected. A finance manager may care most about incorrect allocations; a sales manager may care about promises made without approval. Neither risk is captured by the statement that the new model scored better on a public benchmark.
The UK's voluntary Code of Practice for the Cyber Security of AI, published in January 2025, supplies useful established context rather than a new November legal requirement. Provision 3.1.1 addresses security risks when settings or configuration options change. Its principles also cover human responsibility, asset tracking, documentation, testing and monitoring. A model transition fits naturally into that lifecycle approach. The code is voluntary; following this article is not a certification or a substitute for applicable obligations.
Before switching, agree what would stop the rollout. Examples might include an unauthorised external action, disclosure of restricted information, invalid records reaching the receiving system, or a backlog growing beyond the team's ability to process it. Choose thresholds that match your workflow and consequences. Avoid a generic acceptable error percentage that treats a harmless wording change and an incorrect payment as equivalent.
What this means in practice is giving one named person authority to pause the automated step, with a deputy and clear instructions. They need access to the monitoring and continuity process, not necessarily permission to alter model configuration. Record who decides to resume, what evidence they need and when the remaining rollback window ends. If a risk remains unresolved at that point, the choice is between the tested continuity route and explicit acceptance by the appropriate owner. It is not an automatic licence to continue because the calendar has run out.
The business owner accepts the changed behaviour, not just the successful connection.
A better replacement does not remove the need for recovery
The strongest counterargument is reasonable: the replacement model may be better, and maintaining an old route costs time. Most small businesses should not run a permanent parallel estate just to preserve every historical model. That is not the recommendation. The recommendation is to be honest about when the previous route stops being an option and what can replace it if the new route disappoints.
Benchmarks, demonstrations and supplier recommendations help identify candidates. They do not settle your acceptance decision. The October 2026 benchr market update makes the useful distinction between documented model changes and evidence from the actual workload. Its practical advice includes checking API compatibility, structured output, tool calls, cost and latency against the same tasks. It is secondary commentary; use the supplier's own lifecycle page for the retirement deadline.
For a low-consequence writing assistant with mandatory human review, a modest test pack and a clear way to stop using the assistant may be proportionate. A workflow that can change customer records or trigger payments deserves stronger evidence and tighter controls. Repeating difficult cases can reveal inconsistent behaviour, but the right amount of testing depends on risk. There is no universal number of successful prompts that proves a replacement safe.
Nor should the recovery exercise quietly widen the system's authority. If the alternative model cannot perform a task within the existing permissions, do not grant broader access merely to make the test pass. Fix the integration or choose a narrower continuity route. Local deployment is another possible option, not an automatic escape from maintenance, security or compatibility work. The aim is a service that can fail in a controlled way, not an elaborate promise that AI will never fail. A simple, rehearsed pause can be more useful than an impressive fallback nobody has actually used.
A simple, rehearsed pause can be more useful than an impressive fallback nobody has actually used.
Rehearse the day after the escape hatch closes
Run one final scenario before approving the transition: assume the old model has already disappeared and the replacement has just failed. Do not allow the exercise to call the old route. Can the team identify affected work, stop further automated actions and continue the essential service? This is the situation your continuity plan must handle after the rollback expiry date.
Start with one important workflow rather than a company-wide paperwork exercise. Ask the business owner and the person responsible for the integration to complete the recovery record together. Confirm the exact provider notice, the real endpoint in use and any intermediary supplier's commitments. Anthropic says it provides at least 60 days' notice for publicly released model retirements on its covered platforms. That notice is useful, but it is not a guarantee that your business has identified every hidden dependency or reserved staff time to address it.
Schedule the rehearsal early enough to fix what it reveals. Keep a short evidence record: the test date, participants, dependencies checked, observed recovery time, unresolved failures and the decision taken. After cutover, monitor failures and review a sample of completed work. Continue long enough to include the workflow's important cycle, such as month-end processing, rather than assuming that an uneventful first afternoon represents every use case.
Finally, remove expired recovery claims. An outdated document saying that staff can switch back may be worse than no document because it delays the real response. Mark the previous route as unavailable, update staff instructions and keep the independently tested continuity route current. Your business does not need to freeze itself on an ageing model. It needs to avoid confusing a temporary migration option with a durable recovery capability. The most useful question for your next AI review is simple: if we cannot return to the previous model, what exactly keeps this service working?
Rehearse recovery with the old model unavailable, not merely with the new model switched off.
Key Takeaways
- Record the exact model, hosting route and supplier retirement date for every recovery option.
- Separate temporary technical rollback from an independently tested continuity process.
- Test prompts, retrieval, validation and final business records, not just the replacement model response.
- Give a named owner and deputy authority to pause the automated step against agreed stopping rules.
- Rehearse recovery with the old route unavailable, then remove expired rollback instructions.
