Model Drop

Gemini 3 Pro

November 18, 2025 2 min read

Gemini 3 Pro, which Google DeepMind shipped in preview on November 18, 2025, is the first Gemini 3 model, a reasoning and multimodal flagship built for agentic coding.

Model: Gemini 3 Pro (gemini-3-pro-preview)
Maker: Google DeepMind
Ship date: November 18, 2025
Pricing: $2 / $12 per M tokens (prompts up to 200K; $4 / $18 above)
Context: 1M tokens
Where you can use it:
- Gemini app and AI Mode in Google Search
- Google AI Studio, Vertex AI and Gemini CLI
- Google Antigravity, Google's new agentic IDE
When to reach for it:
- Hard reasoning, math and science work
- Reading charts, screens and long videos
- Skip it where a wrong guess costs you
How it compares:
- Leads GPT-5.1, Sonnet 4.5 and Grok 4
- Ties GPT-5.1, Sonnet 4.5 on repo bug fixes
What makes it different:
- Thinking levels replace token budgets
- Launched alongside the Antigravity IDE
Early reactions: The Batch called it the new benchmark leader but flagged heavy token use and a habit of answering wrong when it could have declined.
Read more: A new era of intelligence with Gemini 3, Gemini 3 Pro model card

The benchmarks

Gemini 3 Pro preview at default settings (high thinking); Google compared Gemini 2.5 Pro, Claude Sonnet 4.5 and GPT-5.1 (high reasoning where reported), using self-reported or leaderboard numbers, and SWE-bench scores follow each maker's own scaffolding.

Google DeepMind's numbers

Humanity's Last Exam (no tools) (expert-level questions): 37.5%. +11.0 over GPT-5.1 (26.5%), +23.8 over Claude Sonnet 4.5 (13.7%), +15.9 over Gemini 2.5 Pro (21.6%).
ARC-AGI-2 (novel visual puzzles): 31.1%. +13.5 over GPT-5.1 (17.6%), +17.5 over Claude Sonnet 4.5 (13.6%), +26.2 over Gemini 2.5 Pro (4.9%).
GPQA Diamond (graduate-level science): 91.9%. +3.8 over GPT-5.1 (88.1%), +8.5 over Claude Sonnet 4.5 (83.4%), +5.5 over Gemini 2.5 Pro (86.4%).
MathArena Apex (hardest contest math): 23.4%. +22.4 over GPT-5.1 (1.0%), +21.8 over Claude Sonnet 4.5 (1.6%), +22.9 over Gemini 2.5 Pro (0.5%).
Video-MMMU (learning from video): 87.6%. +7.2 over GPT-5.1 (80.4%), +9.8 over Claude Sonnet 4.5 (77.8%), +4.0 over Gemini 2.5 Pro (83.6%).
Terminal-Bench 2.0 (agentic terminal coding): 54.2%. +6.6 over GPT-5.1 (47.6%), +11.4 over Claude Sonnet 4.5 (42.8%), +21.6 over Gemini 2.5 Pro (32.6%).
SWE-bench Verified (real GitHub fixes): 76.2%. −0.1 under GPT-5.1 (76.3%), −1.0 under Claude Sonnet 4.5 (77.2%), +16.6 over Gemini 2.5 Pro (59.6%).
SimpleQA Verified (factual recall): 72.1%. +37.2 over GPT-5.1 (34.9%), +42.8 over Claude Sonnet 4.5 (29.3%), +17.6 over Gemini 2.5 Pro (54.5%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2, AA estimate): 28.0. +3.3 over GPT-5.1 (high) (24.7), +7.3 over Claude Sonnet 4.5 (thinking) (20.7), +5.5 over Grok 4 (22.5).
ARC-AGI-1 (ARC Prize verified puzzles): 75.0%. +2.2 over GPT-5.1 (high) (72.8%), +11.3 over Claude Sonnet 4.5 (thinking) (63.7%), +8.3 over Grok 4 (66.7%).
LMArena text Elo (launch score, per Google): 1501.