Built-in AI calls get four minutes and a clear timeout

Workflows that send a lot of text to the built-in AI model, such as a long web page to pull every company or price out of, could take longer than the two minutes a call was given. The workflow got a vague "request failed" error, tried again, failed the same way, and each attempt was still billed because the model kept working after the workflow had stopped waiting. A call now has four minutes to answer. One that runs past that is stopped, is not billed, and ends with a timeout that says what to change.

How it works

• Four minutes per call. Each call has four minutes, including any switch to a backup model when the first one is busy.
• A timeout that says why. A call past four minutes answers with a 504 status, explaining that it was stopped and not billed, and that the same request will take as long again. Split a long page or a large batch across several calls, or ask for a shorter answer.
• Giving up stops the work. When a workflow stops waiting on a call, the model stops generating too, and that call is not billed.
• The same for the API. Calls through the developer API follow the same limits and the same timeout answer.

Also shipped

• Agents writing workflows are told about the four-minute limit, so they split large inputs up front instead of retrying them.