Google's flagship reads text, images and video natively, with a million-token window and Deep Think for the hardest reasoning. Try it here, then open the full studio.
Gemini 3 Pro was Google's November 2025 flagship. Its distinguishing features are scale of input and genuinely native multimodal reasoning.
Published figures, including the rows where Gemini 3 Pro trails. Higher is better; a dash means no directly comparable published figure.
| Benchmark | Gemini 3 Pro | GPT-5.2 | Claude Opus 4.6 |
|---|---|---|---|
| MMMU-Pro· multimodal reasoning | 81.0% | — | — |
| Video-MMMU· video understanding | 87.6% | — | — |
| Global PIQA· commonsense, 100 languages | 93.4% | — | — |
| MMMLU· multilingual Q&A | 91.8% | — | — |
| AIME 2025 (no tools)· competition math | 95.0% | — | — |
| GPQA Diamond· graduate science | 91.9% | 93.2% | — |
| SWE-bench Verified· real GitHub issues | 76.2% | 80.0% | 80.8% |
| ARC-AGI-2· abstract reasoning | 31.1% | — | 69.2% |
| Context window· input tokens | 1M | 400K | 1M |
Figures from Google's Gemini 3 launch materials and Vellum's independent benchmark breakdown. Deep Think raises several of these further. The honest picture: Gemini 3 Pro trails on software engineering, where both Claude Opus 4.6 and GPT-5.2 lead SWE-bench Verified.
Feed it an entire codebase, a stack of legal documents or hours of transcripts in one prompt. Recall holds up — 77.0% on MRCR v2 at 128K average context length.
81.0% on MMMU-Pro and 87.6% on Video-MMMU. It reasons across images and video rather than captioning them, which shows on UI screenshots and video lectures.
31.1% on ARC-AGI-2, rising to 45.1% with Deep Think — a large jump over Gemini 2.5 Pro's 4.9% and a sign of real non-verbal problem solving.
An Elo of 2,439 on LiveCodeBench Pro, well ahead of GPT-5.1's 2,243. Its strength is writing novel, efficient code from scratch rather than patching existing repos.
It tops Vending-Bench 2, which simulates running a business for a year, by staying on task and using tools consistently instead of drifting.
93.4% on Global PIQA across 100 languages, suggesting culturally-aware understanding rather than translation dressed up as fluency.
The model is identical. What differs is everything around it.
Gemini 3 Pro has a distinct shape. It isn't trying to win every benchmark — it's built around two things most rivals handle less naturally: very large inputs and genuinely multimodal reasoning.
A million-token window is only useful if the model can find things in it, and this one can: 77.0% on MRCR v2 at 128K average context. In practice that means entire repositories, multi-document legal analysis, or a full research corpus in a single prompt.
This is the clearest advantage. 81.0% on MMMU-Pro and 87.6% on Video-MMMU reflect reasoning across visual material, not describing it. If your work involves UI screenshots, diagrams, charts or video, it's the first model to reach for.
An Elo of 2,439 on LiveCodeBench Pro puts it well ahead on competitive-style problems. Note the contrast with SWE-bench Verified, where it scores 76.2% against Claude Opus 4.6's 80.8%.
Two clear gaps. Software engineering on real repositories: both Opus 4.6 (80.8%) and GPT-5.2 (80.0%) beat its 76.2% on SWE-bench Verified. And abstract reasoning: Opus 4.6 reaches 69.2% on ARC-AGI-2 against Gemini's 31.1%. Pick Gemini for scale and modality; pick Claude for hard engineering.
thinking_level parameter and a Deep Think mode for the hardest problems.