Google DeepMind

Gemini 3.1 Pro

Released Feb 11, 2026 · knowledge cutoff 2025-01

Status: Superseded
Location: United States / United Kingdom
Type: LLM
Modality: Multimodal
Context window: 1M
Max output: 66K
Speed: 103 tok/s · 21s to first answer token
Price ($/MTok in / out): $2 / $12
Cost per task: $0.29
Open weights: No

Benchmarks8 of 10 reported

GPQA Diamond

94.3%

SWE-bench Verified

80.6%

SWE-bench Pro

54.2%

Terminal-Bench 2.1

70.7%

AIME (latest)retired

98.1%

Humanity's Last Exam

44.4%

LMArena Elo

1486

ARC-AGI-2

77.1%

MMLU-Proretired

—

GDPval-AA v2

—

Strengths & weaknesses

Strengths

Strong on genuinely novel problems rather than recalled ones
Coordinates several tools in one turn without losing track of which returned what
Direct and efficient by default — no preamble, no filler

Weaknesses

Wants the question placed after the data in long prompts, or it under-uses the context
Elaborate prompt scaffolding backfires — it over-analyses the prompt itself
Creative prose is competent but flat

Preview Feb 11, 2026; ARC-AGI-2 reported as 77.1 (76.5 in some sources). Knowledge cutoff Jan 2025 per DeepMind model page. AIME 98.1 is vals.ai-measured (third-party); MMLU-Pro not publicly reported. Terminal-Bench 2.1 70.7 is the official tbench.ai leaderboard figure with the Terminus 2 harness (Gemini CLI harness 67.1; vals.ai's own harness 70.79). Speed per Artificial Analysis' Gemini 3.1 Pro Preview page (Google's API), which carries no effort label. SWE-Bench Pro (public) 54.2% is Google's own model-card figure (deepmind.google/models/model-cards/gemini-3-1-pro); Scale's public leaderboard lists "gemini-3.1-pro (thinking)" lower at 46.10%, a harness difference rather than a correction.

For developers

API model strings

Not researched

Licence

Proprietary — weights not released

Retirement

No retirement announced

Lineage

No recorded predecessor