Gemini 3 Pro

Gemini 3 Pro

Quick answer: Gemini 3 Pro is Google DeepMind's most intelligent model, released in preview on 18 November 2025. It is a natively multimodal reasoning model with a 1M-token context window, positioned as Google's best vibe-coding and agentic model. Google reports 91.9% on GPQA Diamond, 76.2% on SWE-bench Verified, and a leading 1501 Elo on LMArena. It is a closed, API-only model.

At a Glance

Where Gemini 3 Pro leads

• Top-tier reasoning — 91.9% on GPQA Diamond and a leading 1501 Elo on LMArena.
• Native multimodality — leading Video-MMMU (87.6%) and MMMU-Pro (81%), with text, image, video, and audio input.
• Long-horizon planning — tops Vending-Bench 2 across a simulated business year; 1M-token context.

Where it lags

• Coding trails the very top — 76.2% on SWE-bench Verified, below Claude Opus 4.6's 81.42%.
• Closed and API-only with no open weights to self-host, and the launch release is a preview.

Best for: multimodal understanding and long-horizon agentic work where video/audio input and a 1M-token window matter.

What Gemini 3 Pro Is

Gemini 3 Pro is the first release in Google's Gemini 3 era — the model Google describes as its most intelligent yet, combining state-of-the-art reasoning, native multimodality, agentic tool use, and a 1M-token context window in a single system. It launched in preview across Google's product stack on day one: the Gemini app, AI Mode in Search, AI Studio, Vertex AI, Gemini CLI, and Google's new agentic development platform, Antigravity. A separate enhanced reasoning mode, Gemini 3 Deep Think, pushes scores higher and is gated behind additional safety review.

Google's pitch centres on three jobs — learn, build, plan. On reasoning it tops LMArena at 1501 Elo and posts PhD-level scores on Humanity's Last Exam (37.5% without tools) and GPQA Diamond (91.9%). On building it leads WebDev Arena at 1487 Elo, scores 54.2% on Terminal-Bench 2.0, and reaches 76.2% on SWE-bench Verified. On planning it tops Vending-Bench 2, maintaining consistent tool use and decision-making across a full simulated year of running a business — a direct signal for long-horizon agent reliability.

For Benchgen, Gemini 3 Pro is an API-benchmarked model with no open weights. A leaderboard Elo tells you it wins head-to-head preference battles, but only a verified, repeated-trial run on a real domain task shows whether that translates into reliable task completion under a consistent rubric. That reliability-under-repetition view is what Benchgen is built to surface.

Specifications

Pricing

Prompts above 200k tokens are billed at $4.00 / $18.00 per 1M (output includes thinking tokens). Source: Gemini API pricing.

Context Window

Gemini 3 Pro has a 1M-token context window — roughly 1,500 pages of text in a single request — and reasons natively across text, images, video, and audio. That is enough to hold a large codebase, a long set of documents, or hours of video and audio at once, which underpins its leading multimodal and long-horizon planning results.

Public Benchmark Scores

Scores above are reported by Google and shown for context. They are not Benchgen measurements. Figures depend on harness, tool, and effort settings — see Google's evaluation methodology for details. Deep Think mode posts higher scores (e.g. GPQA Diamond 93.8%, HLE 41.0% without tools).

Gemini 3 Pro vs Alternatives

Gemini 3 Pro's edge is multimodality and long-horizon planning rather than peak coding: it matches GPT-5 on SWE-bench Verified but trails Claude Opus 4.6, while undercutting Opus on price and matching its 1M-token window. For a builder, it is the obvious pick when a workload involves video, audio, or very long multimodal context. (Rival scores attributed to OpenAI and Anthropic; not Benchgen measurements.)

How Gemini 3 Pro Performs on Real Agent Tasks

Gemini 3 Pro's most agent-relevant result is Vending-Bench 2: Google reports it tops the leaderboard for long-horizon planning, holding consistent tool use and on-task decision-making across a full simulated year of running a business without drifting. Paired with a 54.2% Terminal-Bench 2.0 score and the Antigravity platform — where agents plan, code, and validate end-to-end via browser and terminal — the model is clearly engineered for autonomous, multi-step work rather than single-turn answers.

What those benchmarks cannot tell you is how Gemini 3 Pro behaves on a specific domain workflow, run many times, under a consistent rubric. Vending-Bench measures planning in a synthetic business sim; what agent builders actually need is reliable, repeated completion of the task in front of them — including how often the model fails, and why. That reliability-under-repetition view is what Benchgen is built to surface, and it is the dimension public benchmarks leave open.

Use Gemini 3 Pro via API

from google import genai

client = genai.Client(api_key="YOUR_API_KEY")

response = client.models.generate_content(
    model="gemini-3-pro",
    contents="Summarize this contract clause...",
)
print(response.text)

Related Models & Benchmarks

• Compare: GPT-5 · Claude Opus 4.6 · Grok 4 · Claude Sonnet 4.6
• Benchmarks: GPQA Diamond · SWE-bench Verified

Frequently Asked Questions

Specs and scores sourced from Google's Gemini 3 announcement. Third-party benchmark scores attributed inline to Google. Benchgen scores, where shown, are produced on Benchgen infrastructure. Last updated 2026-06-17.