Promptfoo 0.123.0 switches GPT-5.6 evals to a new API

Promptfoo's 0.123.0 release changes the default endpoint for GPT-5.6 and every model after it, moving them from chat completions to the Responses API.

That matters for anyone running automated evals against those models. If a config parses or asserts on response shape, the endpoint switch can break tests that were passing for the wrong reason. It can also change what those tests check without throwing an error.

The release adds providers too: Claude Fable and Mythos 5.1, GPT-6 Astra, grok-4.6, Gemini 3.8 Flash, and Muse Spark 1.3. MCP tool calls now show up in response metadata.

The model list is easy to scan in a changelog. The endpoint default is what breaks something three weeks from now, after everyone's forgotten they upgraded. If you're on 0.123.0 and running GPT-5.6 evals, diff a test run before you trust the numbers. The release notes don't say whether older configs get a compatibility shim or just start returning different response shapes, so that check is on you.