Name: Gemini 3 Pro
Author: Google

TL;DR

Google DeepMind's flagship reasoning and multimodal model, first to break 1500 Elo on LMArena (1501) at launch
1M-token context window, 91.9% GPQA Diamond, $2.00/M input and $12.00/M output tokens (<=200K context)
Traded blows with Claude Opus 4.5 and GPT-5.1 at launch, then Google retired the API on March 9, 2026 in favor of Gemini 3.1 Pro

Overview

Google DeepMind shipped Gemini 3 Pro on November 18, 2025, and the rollout was about as aggressive as a model launch gets: same-day availability in the Gemini app, AI Mode in Search, Google AI Studio, Vertex AI, the Gemini CLI, and a brand-new agentic IDE called Google Antigravity. CEO Sundar Pichai framed it as a shift in what these systems do with context, not just text.

"AI has evolved from simply reading text and images to reading the room."

Google CEO Sundar Pichai Sundar Pichai announced Gemini 3 as the company's most intelligent model yet, positioning it as a direct answer to OpenAI and Anthropic's own November 2025 releases. Source: Wikimedia Commons

Underneath the marketing line sits a sparse Mixture-of-Experts transformer with native support for text, images, audio, video, and PDF input, plus a 1,048,576-token context window and up to 65,536 tokens of output. Google has never disclosed a parameter count, and outside estimates (commonly cited in the trillion-plus range) are not confirmed by any primary source, so treat them as speculation rather than spec. Within hours of release, Gemini 3 Pro posted a preliminary 1501 Elo on the LMArena text leaderboard, the first model ever to cross 1500 in that ranking.

The competitive picture at launch was messier than the Elo number suggests. Gemini 3 Pro led on graduate-level science reasoning and math, split coding benchmarks with OpenAI's GPT-5.1, and trailed Anthropic's Claude Opus 4.5 on agentic coding tasks. Three months later, Google shipped Gemini 3.1 Pro, which more than doubled Gemini 3 Pro's reasoning score, and by March 9, 2026, Google shut the original model down completely, migrating every API caller to the newer version. This card documents Gemini 3 Pro as the historical release it now is: superseded, but the model that set up the benchmark race that followed. Our full hands-on review covers the vision and spatial reasoning side in more depth.

Key Specifications

Specification	Details
Provider	Google DeepMind
Model Family	Gemini
Architecture	Sparse Mixture-of-Experts Transformer
Parameters	Not disclosed
Context Window	1,048,576 tokens input / 65,536 tokens output
Input Price	$2.00/M tokens (<=200K), $4.00/M tokens (>200K)
Output Price	$12.00/M tokens (<=200K), $18.00/M tokens (>200K)
Release Date	November 18, 2025
Knowledge Cutoff	January 2025
License	Proprietary (API access only)
Input Modalities	Text, images, audio, video, PDF
Output Modality	Text
Status	Deprecated - shut down March 9, 2026, replaced by Gemini 3.1 Pro

Benchmark Performance

Benchmark	Gemini 3 Pro	Claude Opus 4.5	GPT-5.1
ARC-AGI-2 (reasoning)	31.1%	37.6%	17.6%
GPQA Diamond (science)	91.9%	87.0%	88.1%
MMLU-Pro (knowledge)	89.8%	Not disclosed	Not disclosed
SWE-Bench Verified (coding)	76.2%	80.9%	76.3%
Terminal-Bench 2.0 (agentic)	54.2%	59.3%	58.1%*
MMMU-Pro (multimodal)	81.0%	Not disclosed	Not disclosed
Humanity's Last Exam, no tools	37.5%	Not disclosed	Not disclosed
MathArena Apex (competition math)	23.4%	~1-2%	1.0%

*GPT-5.1-Codex-Max, OpenAI's coding-tuned variant, not the base GPT-5.1 model.

Gemini 3 Pro's strongest showing is on knowledge and science benchmarks: 91.9% on GPQA Diamond and 89.8% on MMLU-Pro were both the highest recorded at launch. The MathArena Apex result is the standout - 23.4% against roughly 1% for GPT-5.1 and Claude, a genuine step change rather than a gradual gain, per Google's own launch benchmarks and independent coverage from InfoQ.

The gap runs the other way on abstract reasoning and agentic coding. Claude Opus 4.5 posted 37.6% on ARC-AGI-2 against Gemini's 31.1%, according to the official ARC Prize leaderboard, and also edged ahead on both SWE-Bench Verified and Terminal-Bench 2.0. See our coding benchmarks leaderboard for how these numbers held up against later releases.

ARC-AGI-2 leaderboard chart showing Gemini 3 Pro and Gemini 3 Deep Think scores plotted against cost per task The official ARC Prize chart puts Gemini 3 Pro at 31.1% on ARC-AGI-2 for $0.81 per task, with the Deep Think variant reaching 45.1% at nearly 100x the cost. Source: arcprize.org

Key Capabilities

Native multimodality is the core pitch. Gemini 3 Pro processes images, video, and audio through the same architecture as text rather than bolting on a separate vision encoder, which is reflected in the 81.0% MMMU-Pro and 87.6% Video-MMMU scores Google reported at launch. In practice that means reading circuit diagrams, parsing scanned documents with mixed layouts, and tracking objects across video frames without the accuracy drop that keyframe-sampling approaches show.

The launch also introduced "vibe coding" and Dynamic View, a generative-UI feature in the Gemini app that turns a single prompt into an interactive interface instead of a wall of text.

Gemini app Dynamic View generating an interactive Van Gogh gallery from a text prompt Dynamic View takes a prompt like "create a Van Gogh gallery with life context for each piece" and builds a working interactive page around it, rather than returning plain text. Source: 9to5google.com

Google paired the model with Google Antigravity, an agent-first IDE where Gemini 3 Pro drives multi-step coding tasks across the editor, terminal, and browser simultaneously. It was also the default model behind an early version of Gemini CLI before that tool moved to Gemini 3.1 Pro. Agentic tool use was a real strength at 54.2% on Terminal-Bench 2.0, just not the strongest in its cohort - both Claude Opus 4.5 and GPT-5.1-Codex-Max scored higher on the same benchmark.

Pricing and Availability

Gemini 3 Pro launched with context-tiered pricing, the same structure Google has kept for every Gemini 3 model since:

Tier	Input	Output
Standard (<=200K tokens)	$2.00/M	$12.00/M
Extended (>200K tokens)	$4.00/M	$18.00/M

That undercut both Claude Opus 4.5 and GPT-5.1 on paper, and Batch API access was supported for asynchronous workloads at launch. None of that matters for new integrations today: Google deprecated gemini-3-pro-preview on the Gemini API and Google AI Studio and shut it down entirely on March 9, 2026, per the official Gemini deprecations page, with gemini-3.1-pro-preview as the mandated replacement. Anyone with the model ID hardcoded into production code broke on that date. For current pricing and access, see the Gemini 3.1 Pro card or our guide to choosing an LLM.

Strengths

Leading science and knowledge scores at launch. 91.9% GPQA Diamond and 89.8% MMLU-Pro were both the highest reported by any model in November 2025
Genuine multimodal architecture. Text, image, audio, and video processed natively, with 81.0% on MMMU-Pro and 87.6% on Video-MMMU
Math breakthrough. 23.4% on MathArena Apex was more than 20x anything GPT-5.1 or Claude posted at the time
Aggressive, coordinated launch. Same-day availability across Search, the Gemini app, AI Studio, Vertex AI, Gemini CLI, and the new Antigravity IDE
Competitive pricing. $2.00/$12.00 per million tokens undercut both Claude Opus 4.5 and GPT-5.1 at standard context lengths

Weaknesses

Trailed on abstract reasoning. 31.1% on ARC-AGI-2 against Claude Opus 4.5's 37.6%, a real gap on tasks requiring novel pattern-solving
Behind on agentic coding. 54.2% on Terminal-Bench 2.0 versus 59.3% for Claude Opus 4.5 and 58.1% for GPT-5.1-Codex-Max
Fully deprecated. Google shut down API access on March 9, 2026; the model is no longer usable for new integrations
Short shelf life. Superseded after roughly three months by Gemini 3.1 Pro, which more than doubled the ARC-AGI-2 score
Undisclosed parameter count. No official figure exists, and third-party estimates are unverified

Gemini 3 Pro Review: Google's Vision AI Powerhouse - our hands-on review of vision and spatial reasoning
Gemini 3.1 Pro - the successor model that replaced Gemini 3 Pro in production
GPT-5.1 - OpenAI's competing launch from the same week
Claude Opus 4.5 - Anthropic's release that led on coding and agentic benchmarks
Google Antigravity Review - the agentic IDE built around Gemini 3 Pro at launch
Gemini CLI Review - Google's terminal agent, originally powered by this model
ChatGPT vs Claude vs Gemini - head-to-head assistant comparison
Understanding AI Benchmarks - what ARC-AGI-2, GPQA Diamond, and SWE-Bench actually measure

FAQ

Is Gemini 3 Pro still available?

No. Google deprecated gemini-3-pro-preview on the Gemini API and AI Studio, shutting it down March 9, 2026. Use Gemini 3.1 Pro instead.

How does Gemini 3 Pro compare to Gemini 3.1 Pro?

Gemini 3.1 Pro scores 77.1% on ARC-AGI-2 versus Gemini 3 Pro's 31.1%, more than doubling reasoning performance in one generation.

What made Gemini 3 Pro notable at launch?

It was the first model to break 1500 Elo on the LMArena leaderboard and posted the highest GPQA Diamond and MathArena Apex scores recorded at the time.

Did Gemini 3 Pro beat Claude Opus 4.5?

Not consistently. Gemini led on science and math benchmarks, but Claude Opus 4.5 scored higher on ARC-AGI-2, SWE-Bench Verified, and Terminal-Bench 2.0.