Gemini 3.1 Pro (Preview)

Gemini 3.1 Pro preview is built for tasks where simple answers are not enough. Stronger core reasoning for complex coding, math, and long-context workflows, with multimodal support and a reported 77.1% verified score on ARC-AGI-2. NOTE: Inputs > 200k tokens are charged at 2x input and 1.5x output rates.

Reasoning
Vision
Audio Input
Video Input
Tool Calling
Structured Output

Added Feb 19, 2026

Pricing

Auto routing · per 1M tokens

Input: $2.00
Output: $12.00
Cache read: $0.20

Compare provider prices

Specifications

Context window: 1M
Max output: 65.5K
Avg output (7d): 1.5K tokens; Longer than 86% of models

Benchmarks

Artificial Analysis

LMArena

Vectara

Design Arena

Sourced from Artificial Analysis.

Intelligence Index

29.7

Better than 84% of models compared

Coding Index

68.8

Better than 80% of models compared

Agentic Index

8.2

Better than 41% of models compared

Agentic work

AutomationBench-AA

Workflow automation with guardrail penalties

35.4%

Better than 53% of models compared

Harvey LAB-AA

Legal agentic work criterion pass rate

58.9%

Better than 7% of models compared

AA-Briefcase

Agentic knowledge work (Elo)

456 Elo

Better than 21% of models compared

GDPval-AA v2

Economically valuable tasks (Elo)

794 Elo

Better than 36% of models compared

Document reasoning

GDP.pdf

Professional PDF reasoning: all-pass rate

17.8%

Better than 65% of models compared

AA-LCR v1.1

Long context reasoning with updated grading

82.0%

Better than 92% of models compared

MLCR-AA

Medical long-context reasoning

15.6%

Better than 51% of models compared

Reasoning

HLE

Humanity's Last Exam

47.0%

Better than 94% of models compared

IFBench

Instruction-following benchmark

77.1%

Better than 96% of models compared

CritPt

Research-level physics reasoning

17.7%

Coding

Terminal-Bench v4.0

Practical coding and terminal tasks

4.0%

Better than 54% of models compared

SciCode

Python programming for scientific computing

58.7%

Better than 91% of models compared

Knowledge

AA-Omniscience Accuracy

Proportion of correctly answered questions

54.8%

AA-Omniscience Hallucination Rate

Rate of incorrect answers among non-correct responses

50.9%

Legacy benchmarks

GPQA Diamond (legacy)

Graduate-level scientific reasoning

94.1%

Better than 99% of models compared

Terminal-Bench Hard (legacy)

Agentic coding and terminal use

53.8%

Better than 96% of models compared

T²-Bench Telecom (legacy)

Conversational AI agents in dual-control scenarios

95.6%

Better than 95% of models compared

AA-LCR (unversioned / legacy)

Long context reasoning evaluation

82.0%

Better than 92% of models compared

GDPval-AA (unversioned / legacy)

Economically valuable tasks

14.6%

Last updated Oct 15, 2026