
Goose 1.60.2, published November 20, repairs how its Google integration sends thinking settings to newer Gemini models. The release also adds Gemini 4 Neon to its model catalog. These are client integration changes, not new model releases by Goose.
The upstream fix describes the earlier problem: newer names could miss a generation-specific check, leaving no thinking configuration in the request despite a user changing the effort setting.
Read the value that actually reaches the provider
The versioned formatter now recognizes Gemini major versions 3 and above for the thinking-level path. It maps the effort choice through levels listed for the model, choosing a supported fallback rather than assuming every model accepts the same maximum.
For example, maximum effort can map to xhigh where that level is listed, or to high on a model without it. The release’s tests include that distinction. We inspected the code and test changes; we did not run a live Google request.
Off is an especially important edge case
The formatter sends a minimal level for Off only where that level is supported. Otherwise, it can omit the thinking configuration. An omitted field should not be interpreted as proof that the remote service disables all internal reasoning.
This is a useful operational distinction because the interface label, the serialized request and the provider’s eventual behavior are three different things. A setting can look consistent in a menu while producing a different request for a different model.
Our analysis: integration tests should inspect requests
A check that only confirms a response arrived would not necessarily catch this class of problem. The system might return useful text even when an intended control never crossed the API boundary. Inspecting the structured request is a more direct test of whether the control was applied.
That still leaves a separate question about the effect of the control. A valid effort value does not by itself promise a particular quality improvement, token count, latency or cost. Those outcomes depend on the service and task, and need a different evaluation.
The Goose repository profile links the Apache-2.0 source and current project identity. The open-source client does not make access to a hosted model free or remove provider terms.
We did not install Goose, alter credentials or run the patch’s test suite. This report is a release and source review, including the final patch revision rather than assuming every test proposed earlier in the pull request remained in the release.
EXPLORE THE SOURCE
Projects in this story
GitHub repositories
- aaif-goose/goose ↗View project details
Primary-source review and original editorial analysis. No software installation, model inference, paid service call, security validation or independent benchmark performed.




