Gemini 3 Flash Preview regresses tool use

This is the kind of silent API break that costs agent teams real time, because the surface area looks unchanged while the model’s tool-use behavior shifts underneath it.

• –Google’s docs still describe `ANY` as supporting function-call selection from allowed tools, so this looks like an unintended regression rather than a documented contract change
• –Multi-step agents, tool-chaining loops, and planner/executor setups are the obvious casualties
• –The fact that Gemini 3.1 Pro still works suggests the issue is specific to Gemini 3 Flash Preview, not the client stack or tool schema
• –Builders should add canaries for tool-call count, retry behavior, and fallback routing across model variants
• –Preview models are only safe in production if you assume behavioral drift and instrument for it