Gemini 3.1 Pro: Google's Reasoning Leap Explained

TL;DR

Google released Gemini 3.1 Pro (preview) on February 19, 2026. The key numbers:

• ARC-AGI-2: 77.1% — more than double Gemini 3 Pro (31.1%), beats Opus 4.6 (68.8%) and GPT-5.2 (52.9%)
• GPQA Diamond: 94.3% — leads all models on graduate-level science
• SWE-bench: 80.6% — matches Opus 4.6 (80.8%) on coding
• Price: $2/$12 per M tokens — cheapest frontier model
• 1M token context — unchanged from Gemini 3 Pro
• Leads on 13 of 16 benchmarks evaluated by Google
• Available now in preview: AI Studio, Vertex AI, Gemini CLI, Gemini app

What Google Announced

On February 19, 2026, Google released Gemini 3.1 Pro — the first ".1" increment in their model versioning. It builds on Gemini 3 Pro (November 2025) by integrating techniques from the Gemini 3 Deep Think series into a more accessible, faster model.

Google's blog describes it as designed for "tasks where a simple answer isn't enough" — complex multi-step reasoning, data synthesis, and agentic workflows.

The headline stat: 77.1% on ARC-AGI-2, the benchmark for novel abstract reasoning. That's more than double Gemini 3 Pro's 31.1%, and significantly ahead of both Opus 4.6 (68.8%) and GPT-5.2 (52.9%). VentureBeat calls it "a Deep Think Mini with adjustable reasoning on demand."

Full Benchmark Breakdown

Where Gemini 3.1 Pro Leads (13 of 16 benchmarks)

Where Competitors Still Win

The Reasoning Leap in Context

ARC-AGI-2 measures a model's ability to solve problems it has never seen before — pure abstract reasoning, not pattern matching from training data. Here's how quickly Gemini improved:

Gemini 3.1 Pro jumped from 31.1% to 77.1% in one version — a 148% improvement. This comes from integrating Deep Think's extended reasoning techniques into the base model.

What Changed vs. Gemini 3 Pro

1. Deep Think Integration

Gemini 3 Deep Think was a separate, slower model optimized for extended reasoning. Gemini 3.1 Pro bakes those techniques into the standard model, with adjustable reasoning depth. You get Deep Think-level reasoning without the Deep Think latency for most tasks.

2. Dramatically Better Reasoning

The numbers speak for themselves:

3. Better Agentic Performance

APEX-Agents (33.5%) and MCP Atlas (69.2%) scores show Gemini 3.1 Pro is significantly more capable as an autonomous agent — tool use, multi-step planning, and self-correction are all improved.

4. Maintained Multimodal Strength

Gemini 3.1 Pro retains Gemini's core advantage: native multimodal processing of text, images, audio, and video within a single context. No other frontier model matches this breadth at this price point.

Pricing

Same price as Gemini 3 Pro — a free upgrade:

Comparison with Competitors

Gemini 3.1 Pro is the cheapest frontier model — 33% cheaper than Sonnet 4.6 on input, and 20% cheaper on output.

Cost Per Session (100K in + 20K out)

Additional cost optimization:

• Batch mode: 50% discount ($0.22/session)


• Context caching: Cached input reads cost 10% of base price

Availability

Where to Use It

API Quick Start

import google.generativeai as genai

genai.configure(api_key="YOUR_API_KEY")
model = genai.GenerativeModel("gemini-3.1-pro-preview")

response = model.generate_content("Your prompt here")
print(response.text)

Custom Tools Endpoint

Google also launched a specialized endpoint for better tool performance:

model = genai.GenerativeModel("gemini-3.1-pro-preview-customtools")

Use this endpoint when building agents that rely heavily on function calling and tool use.

What This Means

The Reasoning Race Heats Up

Three frontier models released in 13 days:

• Feb 6: Claude Opus 4.6 (Anthropic)


• Feb 17: Claude Sonnet 4.6 (Anthropic)


• Feb 19: Gemini 3.1 Pro (Google)

Best-in-Class Reasoning at Budget Pricing

Gemini 3.1 Pro's 77.1% ARC-AGI-2 is the highest reasoning score available, at the lowest price ($2/$12). For tasks requiring novel problem-solving, abstract reasoning, or scientific analysis, it's the clear choice.

Coding Parity

With 80.6% on SWE-bench (vs. Opus 4.6's 80.8% and Sonnet 4.6's 79.6%), Gemini 3.1 Pro is now competitive on coding for the first time. Previous Gemini models trailed Claude significantly on this benchmark.

The Missing Piece: Computer Use

Gemini 3.1 Pro doesn't benchmark on OSWorld (computer use). Claude Sonnet 4.6 leads at 72.5% on this capability. If your workflow involves browser automation, form filling, or desktop control, Claude remains the only viable option.

For Developers Building Products

The practical implications:

• Cheapest reasoning: $0.44/session vs $0.60 (Sonnet) vs $0.80 (GPT-5.2)


• Best for scientific/analytical tasks: 94.3% GPQA Diamond is the highest score available


• Competitive on coding: 80.6% SWE-bench closes the gap with Claude


• Multimodal advantage: Native video/audio processing that Claude and GPT don't match


• Preview status: Not yet GA — expect improvements before general availability

• Google Blog: Gemini 3.1 Pro announcement
• Google DeepMind: Gemini 3.1 Pro Model Card
• 9to5Google: Gemini 3.1 Pro for complex problem-solving
• VentureBeat: Gemini 3.1 Pro first impressions
• MarkTechPost: Gemini 3.1 Pro 77.1% ARC-AGI-2
• OfficeChai: Gemini 3.1 Pro Benchmarks
• GitHub Blog: Gemini 3.1 Pro in GitHub Copilot
• The Decoder: Gemini 3.1 Pro reasoning