gemini-4-argon or gemini-4-hydrogen, modelVersion in the native response can differ from the requested model. The new FAQ explains how to treat this as a model-consistency validation exception instead of accepting HTTP 200 as success, and why it does not imply a policy violation or account downgrade.
The recommended approach is to preserve the original answer, record the requested and returned models, responseId, and business task ID, then isolate the task in a needs_rerun list. After the batch finishes, decide whether to run a targeted second pass. Streaming calls need every chunk checked, not only the first or last returned model.
📖 Read the Gemini 4 model-reporting FAQ
← Back to live updates · 📚 Monthly archive