Gemini 3 Pro

// Overview

Gemini 3 Pro is the first flagship model in Google's Gemini 3 generation, released on November 18, 2025. It is the most capable tier of the Gemini 3 family at launch, positioned for complex reasoning, agentic workflows, and coding. Google describes Gemini 3 Pro as a sparse mixture-of-experts transformer with native multimodal support, trained from scratch rather than fine-tuned from the earlier Gemini 2.5 line.

The model accepts text, images, audio, video, and PDFs, with a 1 million-token context window and output up to roughly 64K tokens. Google calls Gemini 3 Pro its best model for multimodal understanding, and it topped the LMArena leaderboard at launch. Two new controls — a thinking_level parameter to dial reasoning effort and a media_resolution parameter for vision detail — let developers trade off latency, cost, and accuracy.

Gemini 3 Pro is available through the Gemini API in Google AI Studio, on Vertex AI, and inside the Gemini app. It launched at $2 per million input tokens and $12 per million output tokens for prompts of 200K tokens or less, with higher rates for longer prompts. It is a proprietary, closed-weights model. Google later shipped Gemini 3.1 Pro and Gemini 3.5 Pro as successors in this line.

// Benchmarks

1. GPQA Diamond91.9%
2. Humanity's Last Exam (no tools)37.5%
3. SWE-bench Verified76.2%
4. Terminal-Bench 2.054.2%
5. MMMU-Pro81%
6. Video-MMMU87.6%
7. SimpleQA Verified72.1%
8. MathArena Apex23.4%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Launch API pricing. Context caching and Batch API reduce costs further.

Pricing source ↗

// Strengths

• State-of-the-art reasoning, scoring 91.9% on GPQA Diamond and 37.5% on Humanity's Last Exam (no tools)
• Strong agentic coding, with 76.2% on SWE-bench Verified and 54.2% on Terminal-Bench 2.0
• Best-in-class multimodal understanding across image, video, audio, and PDF inputs (81% MMMU-Pro, 87.6% Video-MMMU)
• 1M-token context window for reasoning over long documents, codebases, and mixed media
• Developer controls for reasoning depth (thinking_level) and vision detail (media_resolution) to tune cost and latency
• Topped the LMArena leaderboard at launch for general capability

// Best for

• Complex multi-step reasoning and analysis over long documents or research material
• Agentic coding tasks: bug fixing, repo-scale changes, and terminal-driven workflows
• Multimodal understanding combining text, images, video, audio, and PDFs
• Building autonomous agents and tool-using assistants via the Gemini API or Vertex AI
• Long-context tasks such as summarizing or querying entire codebases and large document sets
• Factual question answering with reduced hallucination (72.1% on SimpleQA Verified)

// How to access

// Gemini Pro — every version

The full lineage of the Gemini Pro line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.