Gemini 3.1 Pro Preview Custom Tools

This variant of the flagship Gemini 3.1 Pro positions itself as an optimization for developers building coding agents and complex workflow systems. The core innovation addresses a specific friction point in production AI systems: models that default to a catch-all bash tool instead of selecting the most efficient option from available integrations. By refining tool selection behavior, the model demonstrates improved function calling reliability, ensuring it picks targeted third-party or user-defined functions over generic alternatives when the latter would be slower or less appropriate. It maintains the full multimodal reasoning span of its base—processing text, images, video, audio, and code within the same pipeline—while the 1M-token context window enables deep reasoning across extensive codebases or documentation.

The model builds on the base architecture's strengths in software engineering, with benchmark scores including 94.1% on GPQA and 58.9% on SciCode, demonstrating solid performance across academic and applied coding challenges. The tool-selection tuning particularly shines in multi-tool workflows where developers chain multiple function calls together; the reliability gains mean fewer fallback scenarios and more predictable agentic behavior. Available through Google Gemini, Vertex AI, and OpenRouter, the preview endpoint gives teams early access to the refined function-calling behavior while the underlying multimodal reasoning remains intact. For teams building AI-powered coding assistants, automated testing pipelines, or complex agentic workflows that depend on precise tool selection, this variant offers a purpose-built alternative to the standard preview.