Help for owners
My AI automation stopped working after the AI model changed
Short answer
AI providers retire old models and publish new ones, and a workflow tied to one model can fail or change output when that happens. Retired models stop accepting requests, and a replacement may format or word its answers differently. Find which model you use, check the provider's notice, and test before switching.
Does this sound like your situation?
- The automation worked for months and then started failing.
- It still runs but the output looks different, shorter, longer or in another layout.
- A later step complains it cannot read the AI's reply.
- The provider emailed me about a model being retired and I did not act on it.
- Errors appeared the day a model name changed.
Why does this happen?
The model you call was retired
OpenAI gives each model a state of Active, Deprecated or Shutdown. A deprecated model works until its shutdown date and then no longer accepts standard API requests. OpenAI says affected customers are notified by email and in its documentation, and each entry names a recommended replacement.
Confirmed by developers.openai.com
Anthropic follows the same pattern
Its documentation says retired models fail when called, that customers with active deployments are notified, and that it gives at least 60 days' notice before retiring a publicly released model.
Confirmed by platform.claude.com
A new model words things differently
Even a recommended replacement can answer in another style or layout. If the next step expects a fixed format, such as a number on its own line or a particular label, a small difference breaks it.
Technical detail: The model's reply breaks JSON.parse
Request settings are removed or renamed
Providers also remove request parameters. Anthropic's page lists an example of a setting that returns an error on newer API versions. A workflow that sends an old setting can fail even though the model name is right.
Confirmed by platform.claude.com
The tool in the middle changed
If the automation runs inside a no-code tool, the tool itself may have swapped the default model or retired a connector, without you touching anything.
What can I do about it today?
- Find the first failed runOpen the run history of your automation and note the date and the exact error message of the first failure. Copy it into a note.
- Search your email for the provider's noticeLook for messages from the AI provider about deprecation or retirement, including spam, and in the inbox of whoever created the account.
- Write down which model it callsThe model name is in the automation's settings. Do not change it yet; just record it.
- Pause what could do harmIf the automation sends emails, creates invoices or updates records, pause that step so wrong output does not spread. Pausing is reversible.
- Do not just swap the model name in productionTry the replacement on a copy or on a few sample items first, and compare the results with the old ones.
When should I bring someone in?
- The workflow touches customers or money and wrong output is already going out.
- Nobody knows who built it or where the keys are.
- The replacement model gives answers the next step cannot handle.
How do I handle it?
- Reproduce the failureI take real inputs from the failed runs and call the old and new model with them, to see exactly what changed.
- Pin down the output formatWhere a later step needs a fixed shape, the output is requested and checked against a defined structure, with a clear path for replies that do not fit.
- Compare before switchingA set of real examples is run through both models and compared on correctness, as OpenAI also advises, with cost and speed noted.
- Switch with a rollbackThe new model is turned on in a controlled way, with the previous setting kept so it can be restored.
- Put the date on a calendarRetirement notices go to a shared address and each known date is written down with an owner.
The price depends on what I find, so I quote in writing after a short discovery call. Tell me what is happening and I reply within 24 hours.
What do owners ask about this?
Will it keep breaking every time a model is retired?
Retirements are routine, so the aim is to make switching a planned, tested step instead of an emergency.
Can I stay on the old model?
Only until its shutdown date. After that, requests to a retired model fail, according to both providers' documentation.
Is the new model always better?
Not for every task. That is why a sample of your own cases should be compared before switching.
Sources
How was this page made?
I drafted this page with AI assistance from the official documents listed under Sources. Found a mistake? Tell me.
For your developer: the technical fixes
SyntaxError: Unexpected token '`', "```json ..." is not valid JSONThe model's reply breaks JSON.parseInvalid parameter: 'response_format' of type 'json_schema' is not supported with model version `gpt-4-turbo`OpenAI rejects response_format json_schema for the model you choseYou tried to access openai.ChatCompletion, but this is no longer supported in openai>=1.0.0Old openai.ChatCompletion or openai.Embedding code breaks on a new SDK