Model Drop

Gemini 3.1 Pro

February 19, 2026 2 min read

Gemini 3.1 Pro, which Google DeepMind shipped in preview on February 19, 2026, is a mid-cycle upgrade to Gemini 3 Pro with much stronger core reasoning.

Model: Gemini 3.1 Pro (gemini-3.1-pro-preview)
Maker: Google DeepMind
Ship date: February 19, 2026
Pricing: $2 / $12 per M tokens (prompts up to 200K; $4 / $18 above)
Context: 1M tokens
Where you can use it:
- Gemini app, plus NotebookLM on AI Pro/Ultra
- Google AI Studio, Gemini CLI, Android Studio
- Antigravity, Vertex AI and Gemini Enterprise
When to reach for it:
- Novel logic puzzles and abstract reasoning
- Agentic web research and MCP tool chains
- Skip it for expert office deliverables
How it compares:
- Matches Claude Opus 4.6 on repo bug fixes
- Leads GPT-5.2 on reasoning and search
What makes it different:
- Generates animated SVGs straight from text
- Separate endpoint tuned for custom tools
Early reactions: Hacker News users praised its reasoning and its price next to Claude Opus 4.6, but several said it still stumbles at finishing agentic coding tasks.
Read more: Gemini 3.1 Pro: A smarter model for your most complex tasks, Gemini 3.1 Pro model card

The benchmarks

Gemini 3.1 Pro at Thinking (High); Google picked Claude Opus 4.6 and Claude Sonnet 4.6 at max thinking and GPT-5.2 and GPT-5.3-Codex at xhigh, with rival scores mostly self-reported, Terminal-Bench on the Terminus-2 harness, and SWE-bench following each maker's own scaffolding.

Google DeepMind's numbers

Humanity's Last Exam (no tools) (expert-level questions): 44.4%. +4.4 over Claude Opus 4.6 (40.0%), +11.2 over Claude Sonnet 4.6 (33.2%), +9.9 over GPT-5.2 (34.5%).
ARC-AGI-2 (novel visual puzzles): 77.1%. +8.3 over Claude Opus 4.6 (68.8%), +18.8 over Claude Sonnet 4.6 (58.3%), +24.2 over GPT-5.2 (52.9%).
GPQA Diamond (graduate-level science): 94.3%. +3.0 over Claude Opus 4.6 (91.3%), +4.4 over Claude Sonnet 4.6 (89.9%), +1.9 over GPT-5.2 (92.4%).
Terminal-Bench 2.0 (agentic terminal coding): 68.5%. +3.1 over Claude Opus 4.6 (65.4%), +14.5 over GPT-5.2 (54.0%), +3.8 over GPT-5.3-Codex (64.7%).
SWE-bench Verified (real GitHub fixes): 80.6%. −0.2 under Claude Opus 4.6 (80.8%), +1.0 over Claude Sonnet 4.6 (79.6%), +0.6 over GPT-5.2 (80.0%).
SWE-bench Pro (harder agentic coding): 54.2%. −1.4 under GPT-5.2 (55.6%), −2.6 under GPT-5.3-Codex (56.8%).
BrowseComp (agentic web search): 85.9%. +1.9 over Claude Opus 4.6 (84.0%), +11.2 over Claude Sonnet 4.6 (74.7%), +20.1 over GPT-5.2 (65.8%).
GDPval-AA (expert office tasks, Elo): 1317. −289 under Claude Opus 4.6 (1606), −316 under Claude Sonnet 4.6 (1633), −145 under GPT-5.2 (1462).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2): 29.7. −2.2 under Claude Opus 4.6 (max) (31.9), −0.7 under GPT-5.2 (xhigh) (30.4).
ARC-AGI-1 (ARC Prize verified puzzles): 98.0%. +4.0 over Claude Opus 4.6 (high) (94.0%), +11.8 over GPT-5.2 (xhigh) (86.2%).