Gemini 3.1 Pro enters API preview

Google is signaling that frontier model competition is shifting from raw benchmark bragging to dependable execution inside developer tooling.

• –The official docs frame Gemini 3.1 Pro Preview around software engineering behavior, precise tool usage, and reliable multi-step execution, which is exactly where coding agents succeed or fail
• –A 1,048,576-token input window and support for code execution, function calling, structured outputs, URL context, and search grounding make it better suited for full-repo tasks than chat-only workflows
• –The separate `-customtools` endpoint is especially notable because Google is optimizing for agent stacks that mix bash and custom tools, not just generic prompting
• –The surrounding community conversation is already treating it as a direct coding rival to Claude Opus and GPT Codex-class models, which raises the pressure on real-world evals over synthetic scores