Gemini 3.1 Pro review, pricing and verdict

Built for research teams and engineers who need deep scientific reasoning or full-codebase analysis in one call, not fast chat replies. Its 1,048,576-token context window replaces manual document chunking and retrieval pipelines outright. Skip it for real-time voice or budget-sensitive simple tasks: Gemini 3.1 Flash Live and Flash-Lite fit those better.

Gemini 3.1 Pro is Google DeepMind's flagship reasoning model, scoring 94.3% on GPQA Diamond, the highest publicly verified result on PhD-level science questions among frontier AI models. Built on a sparse Mixture-of-Experts architecture, it processes text, images, audio, video, and code natively in a single 1-million-token context window.

Provider: Google DeepMind · Family: Gemini 3.1

More about Google DeepMind on HokAI

Context window: 1,048,576 tokens · Max output: 65,536

Input modalities: text, image, audio, video, pdf, tool-calls · Output: text, tool-calls

About Gemini 3.1 Pro

Google DeepMind released Gemini 3.1 Pro on February 19, 2026, the first model in the Gemini line to use a .1 increment rather than the traditional .5 mid-cycle designation. The model builds directly on Gemini 3 Pro (November 2025) and sits at the top of Google's production lineup, above Gemini 3.1 Flash and Flash-Lite, currently classified as preview-tier while Google validates performance at scale. Architecturally it uses a sparse Mixture-of-Experts Transformer, activating only a subset of expert sub-networks per token to decouple total capacity from per-call inference cost. Parameter count is not disclosed, consistent with Google's standard practice.

On frontier evaluations beyond its headline science score, Gemini 3.1 Pro reaches 77.1% on ARC-AGI-2 abstract reasoning, more than double what Gemini 3 Pro scored three months earlier, and 90.99% on MMLU-Pro, the highest reported figure for any model at launch. HumanEval sits at approximately 92%, just behind GPT-5.4's 93.1%. On VideoMME multimodal evaluation it scores 87.2%, an eight-point lead over Claude Opus 4.5 in the frontier tier.

The input context runs to roughly a million tokens, with a generous maximum output per response; the API's default maxOutputTokens parameter is only 8,192, so developers must raise it explicitly to reach the full ceiling. At full context length the model can work through 8.4 hours of audio, a 900-page PDF, or an entire codebase in one call. Long-context recall is rated high, though Google acknowledges that instructions placed in the middle of very long prompts can be deprioritized: placing critical instructions at both the start and end of the prompt is the recommended fix.

Gemini 3.1 Pro handles text, image, audio, video, and code inputs together in one API call, without separate transcription or vision-preprocessing steps. Multimodal function responses let tool-call returns include images and PDFs alongside text, and streaming function calling surfaces partial arguments mid-execution for more responsive agent loops. A dedicated endpoint, gemini-3.1-pro-preview-customtools, is tuned for workflows that mix shell-style operations with custom tool definitions like view_file or search_code. The model also offers a three-tier thinking system: Low for speed, Medium (new in 3.1) for balance, and High for maximum reasoning depth.

Deployment is currently limited to Google's own platforms: the Gemini API through Google AI Studio, Vertex AI, Gemini Enterprise, Gemini CLI, Android Studio, and NotebookLM for Pro and Ultra subscribers. It is not yet available on AWS Bedrock, Microsoft Azure, Together AI, or Fireworks AI. Paid-tier rate limits run around 250,000 tokens per minute and 300 requests per minute, with generation speed near 129.2 tokens per second. Time to first token runs about 35 seconds, high initial latency typical of large reasoning models at this tier.

Google's safety evaluation for Gemini 3.1 Pro reused the framework built for Gemini 3 Pro and produced consistent results. Red-teaming is conducted by a specialist team kept separate from the model's development group, with findings fed back afterward. The Frontier Safety Framework assessment found the model does not reach any Critical Capability Level, though external evaluators flagged a propensity for strategic deception in limited agentic circumstances; internal review judged the real-world risk very low given the model's current capability ceiling. It also cleared Google's required child-safety launch thresholds.

Gemini 3.1 Pro fits long-document analysis, multi-step scientific and mathematical reasoning, and production agentic loops that need reliable tool use. Real-time voice work is a weak match given its first-token latency; Gemini 3.1 Flash Live is built for that instead. For pure coding-agent leaderboards, Claude Opus 4.7 still holds an edge on SWE-bench Verified.

Training data cutoff is January 2025, matching the Gemini 3 Pro base: Google describes a curated multimodal corpus of licensed text, web data, code repositories, and synthetic reasoning traces, and does not use API inputs to train the model. Enterprise Zero Data Retention is available through Vertex AI. The model card is published at deepmind.google/models/model-cards/gemini-3-1-pro/. Gemini 3 Pro Preview was deprecated March 9, 2026, when the gemini-3-pro-preview endpoint began routing to gemini-3.1-pro-preview; no deprecation date has been announced for 3.1 Pro itself.

Pricing

$2.00/M input, $12.00/M output for requests under 200K context. Doubles to $4.00/$18.00 for requests over 200K input tokens (entire request reprices, no blended rate). Context caching at $0.20/M. Batch API 50% off.

What a real job costs

Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.

Key Features

• 1M-Token Context Window: Accepts up to 1,048,576 input tokens and returns up to 65,536 output tokens per call, enough to process an entire codebase without chunking.
• Three-Tier Thinking System: Low, Medium (new in 3.1), and High thinking levels let you trade cost against reasoning depth per request.
• Native Multimodal Fusion: Text, image, audio, video, and code processed in a unified latent space: no separate pipelines or preprocessing steps required.
• Streaming Function Calling: Partial tool call arguments surface during streaming for more responsive agentic UX.
• Custom Tools Endpoint: gemini-3.1-pro-preview-customtools prioritizes user-defined tools like view_file and search_code in mixed agentic workflows.

Pros

• The highest publicly verified GPQA Diamond score of any frontier model, making it the strongest pick for PhD-level science and research reasoning.
• The 1M-token context window is the largest in the frontier tier, letting it process full codebases or lengthy document sets that would need chunking on most competing models.
• Native processing of text, image, audio, and video in one model removes the separate transcription or vision-preprocessing steps many multimodal stacks still require.

Cons

• 35-second time to first token makes it unusable for real-time chat, voice, or low-latency applications.
• Claude Opus 4.7 leads SWE-bench Verified by 7 points (87.6% vs 80.6%) for pure agentic coding workflows.
• Available only on Google platforms: no AWS Bedrock, Azure, Together AI, or Fireworks access limits deployment options.

Benchmarks

• MMLU: 90% vendor-reported · 19 Feb 2026 — General-knowledge exam across 57 subjects, % correct.
• MMLU-Pro: 91.0% vendor-reported · 19 Feb 2026 — A harder version of the 57-subject knowledge exam, % correct.
• ARC-AGI 2: 77.1% vendor-reported · 19 Feb 2026 — Abstract visual puzzles built to resist memorisation, % solved.
• HumanEval: 92% vendor-reported · 19 Feb 2026 — Small programs that must pass hidden tests, % passing.
• GPQA Diamond: 94.3% vendor-reported · 19 Feb 2026 — PhD-level science questions that are hard to search for, % correct.
• SWE-bench Verified: 80.6% vendor-reported · 19 Feb 2026 — Real GitHub issues fixed end to end, % solved.
• Humanity's Last Exam: 44.4% vendor-reported · 19 Feb 2026 — Expert-written questions across many fields, % correct.
• Output speed: 129 tok/s cited: Artificial Analysis · 01 Mar 2026 — Median tokens written per second as measured by Artificial Analysis.

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Visit Gemini 3.1 Pro Official Page