Models

Gemini 3.1 Pro

Google DeepMind's Gemini 3.1 Pro leads on 13 of 16 benchmarks with 77.1% ARC-AGI-2, 94.3% GPQA Diamond, and a 1M-token context window at $2/M input.

James Kowalski By James Kowalski • • 7 min read
Gemini 3.1 Pro

Overview

Google DeepMind released Gemini 3.1 Pro on February 19, 2026, and the numbers back up the hype. The model posted leading scores on 13 of 16 benchmarks evaluated by Google, including a 77.1% on ARC-AGI-2 - more than double the score of its predecessor Gemini 3 Pro. That single number tells you this is not a minor point release. It is a generational leap in reasoning capability.

Built on a Transformer-based Mixture-of-Experts architecture, Gemini 3.1 Pro processes text, images, audio, video, and code across a 1M-token context window with up to 64K tokens of output. It ships with dynamic thinking by default, meaning the model allocates more internal reasoning to harder problems rather than producing immediate responses. This puts it squarely in the "thinking model" category alongside Claude Opus 4.6 and GPT-5.2, trading latency for accuracy on complex tasks.

The competitive picture is tight at the top. On Chatbot Arena, Gemini 3.1 Pro sits at roughly 1500 Elo versus Claude Opus 4.6's 1504 - essentially tied in human preference voting. But the benchmark story diverges sharply depending on the task category. Gemini dominates on reasoning and scientific knowledge. Claude holds the edge on software engineering and economically valuable enterprise tasks. Picking between them now requires knowing exactly what you need the model for, which is a healthy sign for the field. See our overall LLM rankings for the full picture.

Key Specifications

SpecificationDetails
ProviderGoogle DeepMind
Model FamilyGemini
ArchitectureTransformer-based Mixture of Experts
ParametersNot disclosed
Context Window1,000,000 tokens input / 64,000 tokens output
Input Price$2.00/M tokens (<=200K), $4.00/M tokens (>200K)
Output Price$12.00/M tokens (<=200K), $18.00/M tokens (>200K)
Release DateFebruary 19, 2026
LicenseProprietary (API access only)
Input ModalitiesText, images, audio, video
Output ModalityText
Dynamic ThinkingEnabled by default

Benchmark Performance

BenchmarkGemini 3.1 ProClaude Opus 4.6GPT-5.2
ARC-AGI-2 (reasoning)77.1%68.8%52.9%
GPQA Diamond (science)94.3%91.3%92.4%
MMLU (knowledge)92.6%91.1%89.6%
HLE - no tools (reasoning)44.4%40.0%34.5%
SWE-Bench Verified (coding)80.6%80.8%80.0%
Terminal-Bench 2.0 (agentic)68.5%65.4%54.0%
AIME 2025 (math)100%--
LiveCodeBench Pro (Elo)2887--
GDPval-AA (enterprise)131716061462
MRCR v2 128K (long context)84.9%--

The headline stat is ARC-AGI-2 at 77.1%. This benchmark tests a model's ability to solve entirely novel logic patterns it has never seen, and Gemini 3.1 Pro's score represents a 2.5x improvement over its predecessor. On GPQA Diamond, a graduate-level science reasoning test, it hits 94.3% - the highest recorded score. It matches or exceeds competitors on most knowledge and reasoning benchmarks.

The cracks show on enterprise tasks. The GDPval-AA benchmark measures performance on economically valuable work like financial modeling, strategic planning, and business documentation. Gemini 3.1 Pro's Elo of 1317 trails Claude Opus 4.6 (1606) by a wide margin, and even Claude Sonnet 4.6 (1633) outscores it decisively. If your workload is business-critical document synthesis, the benchmarks point you toward Claude. For a deeper dive into what these numbers actually measure, see our guide to understanding AI benchmarks.

Key Capabilities

Gemini 3.1 Pro's strongest suit is deep reasoning across multiple modalities. It can ingest entire code repositories, lengthy research papers, hours of video, and complex multi-source datasets within its 1M-token context window, then reason across all of them simultaneously. The dynamic thinking system lets you tune reasoning depth via the thinking_level parameter, balancing quality against latency and cost depending on your use case.

Agentic performance is a highlight. On Terminal-Bench 2.0, which evaluates multi-step tool-use workflows, Gemini 3.1 Pro scored 68.5% - above both Claude Opus 4.6 (65.4%) and GPT-5.2 (54.0%). It also scored 69.2% on MCP Atlas, a tool coordination benchmark. Google has been pushing agentic capabilities hard, and the numbers suggest it is paying off. The model is available in GitHub Copilot and Google's own Gemini CLI for coding workflows.

Scientific and mathematical reasoning is where this model genuinely separates from the pack. A perfect 100% on AIME 2025 with code execution, 94.3% on GPQA Diamond, and 44.4% on HLE (a test designed to be too hard for current models) paint a clear picture: if your use case involves research synthesis, complex math, or domain-expert reasoning, Gemini 3.1 Pro is the current leader.

Pricing and Availability

Gemini 3.1 Pro uses tiered pricing based on prompt length:

TierInputOutput
Standard (<=200K tokens)$2.00/M$12.00/M
Extended (>200K tokens)$4.00/M$18.00/M
Batch API (50% discount)$1.00/M$6.00/M
Context Caching$0.20/M-

At the standard tier, Gemini 3.1 Pro is significantly cheaper than Claude Opus 4.6 ($15.00/$75.00 per M tokens) and GPT-5.2 ($10.00/$30.00 per M tokens). The Batch API halves costs for asynchronous workloads, and context caching can reduce repeated input costs by up to 75%. For cost-conscious teams, this pricing makes Gemini 3.1 Pro the best value among frontier reasoning models. Check our cost efficiency leaderboard for full comparisons.

The model is available in preview through the Gemini API via AI Studio, Vertex AI, Gemini Enterprise, Gemini CLI, Android Studio, and NotebookLM. It is also in public preview in GitHub Copilot. Google Gemini app users with AI Pro and Ultra plans get higher usage limits. There is no free tier for API access.

Note: "preview" status means Google is still validating agentic workflows before general availability. Developers have reported capacity issues including 99-hour lockouts and phantom quota drain during the preview period, so plan accordingly if you are building production systems.

Strengths

  • Dominant reasoning benchmarks. 77.1% ARC-AGI-2 and 94.3% GPQA Diamond are the highest scores recorded by any model, with clear leads over Claude Opus 4.6 and GPT-5.2
  • Best price-to-performance ratio. At $2/$12 per M tokens, it undercuts both Opus and GPT-5.2 while matching or exceeding them on most benchmarks
  • 1M-token context window. Processes entire codebases, research paper collections, and long video in a single pass
  • Strong agentic tool use. Leading scores on Terminal-Bench 2.0 and MCP Atlas for multi-step autonomous workflows
  • Multimodal native. Text, image, audio, and video inputs processed natively, not bolted on

Weaknesses

  • Enterprise task performance. GDPval-AA score of 1317 trails Claude significantly (1606), suggesting weakness on business-critical document synthesis and strategic analysis
  • Long-context degradation. Reports of accuracy degradation beyond 120-150K tokens, with fabricated details appearing at extreme context lengths despite the 1M-token window
  • Preview stability issues. Capacity constraints, quota bugs, and broken tool calling reported by developers during the preview rollout
  • Structured output inconsistency. Only 84% schema-valid JSON responses without retries in testing, problematic for production pipelines requiring strict output formats
  • Latency at scale. Time to first token averaging 29 seconds, and multi-file queries with images taking 7-9 seconds per response via API

Sources

Tagged
Gemini Google Google DeepMind Reasoning Multimodal 1M Context Mixture of Experts
Gemini 3.1 Pro
About the author AI Benchmarks & Tools Analyst

James is a software engineer turned tech writer who spent six years building backend systems at a fintech startup in Chicago before pivoting to full-time analysis of AI tools and infrastructure.