What
When Claude Code is set up to route requests through upstream providers (the services that actually run the models), each provider can list the models it accepts. Claude Code checks those lists and warns about problems. These checks used to run only for Mantle. Now:
- They run for every provider.
- Wildcard entries are understood.
- An entry that matches no model that can be sent to that provider gets a warning telling you to check the spelling.
The warning about the 1-million-token context window for claude-sonnet-4-5, when it is pulled in through auto_include_builtin_models, now appears only when a Vertex provider in your setup can actually serve that model.
Why
You get fewer warnings that don't apply to you, and misspelled model names are caught for providers other than Mantle too.