← Back to directory
Google DeepMind
Gemini 3 Pro
Released Nov 18, 2025 · knowledge cutoff 2025-01
- Status
- Superseded
- Location
- United States / United Kingdom
- Type
- LLM
- Modality
- Multimodal
- Context window
- 1M
- Max output
- 66K
- Speed
- —
- Price ($/MTok in / out)
- $2 / $12
- Cost per task
- —
- Open weights
- No
Benchmarks9 of 10 reported
MMLU-Proretired
89.8%
GPQA Diamond
91.9%
SWE-bench Verified
76.2%
SWE-bench Pro
43.3%
Terminal-Bench 2.1
73.9%
AIME (latest)retired
95%
Humanity's Last Exam
37.5%
LMArena Elo
1501
ARC-AGI-2
31.1%
GDPval-AA v2
—
Strengths & weaknesses
Strengths
- Strong on abstract, unfamiliar problems rather than remembered ones
- Terse and direct by default — it answers instead of performing
- Deep Think mode for the rare problem worth minutes of compute
Weaknesses
- The terseness reads as cold and needs explicit prompting to open up
- Old prompt-engineering habits — elaborate personas, hand-written chain-of-thought — make it worse, not better
- Deep Think sits behind the top consumer tier rather than the API
SWE-Bench Pro (public) 43.3% is Google's own figure (cited in the Gemini 3.1 Pro model card's comparison column) and matches Scale's public leaderboard entry for gemini-3-pro-preview (43.30%). Terminal-Bench 2.1 73.9% (±1.3) is the Terminus 2 harness entry dated 2025-11-18 on the tbench.ai leaderboard as mirrored by Snorkel AI — third-party; the Gemini CLI entry scores 65.8%.
For developers
API model strings
- google
gemini-3-pro-preview
Licence
Proprietary — weights not released
Retirement
No retirement announced
Lineage
No recorded predecessor