Gemini 3.1 Pro Preview

Overview. Gemini 3.1 Pro Preview is Google's frontier reasoning model, released in public preview on February 19, 2026. It combines mandatory chain-of-thought reasoning with a 1,048,576-token context window, 65,536-token max output, and multimodal input across text, image, audio, video, and file formats. Developers reach for it when a task needs deliberate, multi-step reasoning over very large or mixed-format input rather than a fast single-pass answer.

Key strengths. Chain-of-thought is mandatory rather than optional on this model, with configurable reasoning effort levels (high, medium, low), which suits tasks that benefit from deeper deliberation over a quick response. Its context window is among the largest available for a hosted model, letting it hold entire document sets, long transcripts, or multi-file codebases in a single call without chunking. Prior LMSys Arena evaluation placed it solidly within frontier-class territory, though the most recent leaderboard snapshot could not be freshly re-verified this cycle.

When to use it. Use this model for long-horizon research, document analysis, or writing tasks where reasoning over very large or multi-file input matters more than raw response speed. It also accepts native image, audio, and file input, making it a fit for multimodal document and media understanding. It is not Google's dedicated choice for heavy agentic tool-calling workloads — a separate custom-tools-tuned sibling variant exists specifically for that use case.

Limitations. Chain-of-thought reasoning is always-on and cannot be disabled, which adds latency relative to non-reasoning models and makes this a poor fit for simple, low-latency lookups. Output is text-only, with no native audio or image generation. Its knowledge cutoff is January 2025, and independent third-party benchmark placements for this model have shown notable movement across recent evaluation snapshots, which we flag for transparency rather than as a settled capability signal.

Sovereignty note. Gemini 3.1 Pro Preview is built by Google, a US-headquartered creator, and served under Google's proprietary Gemini API terms. Our research did not surface an explicit training-on-input opt-out or EU data-residency guarantee for this model, so it does not carry Stav's Sovereign designation.