AI

Google Releases Gemini 3.1 Pro with 2× Reasoning Performance

On February 19, 2026, Google DeepMind released Gemini 3.1 Pro — its most capable model to date and the first in the Gemini family to receive a “.1” mid-cycle update rather than the usual “.5” increment. The designation signals a focused intelligence upgrade: dramatic reasoning gains and stronger agentic performance, without a broad feature overhaul.

Abstract visualization of a glowing neural network sphere representing Gemini 3.1 Pro's advanced reasoning capabilities
Illustration generated by AI

What’s New in 3.1 Pro

The headline improvement is reasoning. Gemini 3.1 Pro scores 77.1% on ARC-AGI-2, the benchmark designed to test a model’s ability to solve entirely novel logic patterns that cannot be memorized from training data. That figure is more than double the score achieved by Gemini 3 Pro — the largest single-generation reasoning jump seen in any frontier model family so far.

Other benchmark highlights include:

  • GPQA Diamond: 94.3% — graduate-level scientific reasoning
  • SWE-Bench Verified: 80.6% — autonomous software engineering tasks
  • LiveCodeBench Pro: 2887 Elo — competitive programming performance
  • APEX-Agents — agentic task scores nearly doubled vs. Gemini 3 Pro
  • BrowseComp: 85.9% — autonomous web research

Google attributes these gains to integrating reasoning advances that originally debuted in the experimental Gemini 3 Deep Think mode, now baked into the standard model for all users.

Technical Specifications

Gemini 3.1 Pro is a natively multimodal model built on a Transformer-based Mixture-of-Experts architecture. It processes text, images, audio, video, and entire code repositories within a 1 million token context window, and can output up to 64,000 tokens in a single response — a significant leap that benefits long-form code generation, document synthesis, and extended agentic workflows.

A key developer-facing addition is the thinking_level parameter, which now includes a Medium option alongside the existing Low and High settings. This lets developers tune the trade-off between reasoning depth, output latency, and API cost within a single call, without switching between model variants.

Pricing remains unchanged from Gemini 3 Pro at $2 / $12 per million input/output tokens — considerably more affordable than comparable frontier models at similar performance tiers.

Agentic and Developer Access

Gemini 3.1 Pro was built with agentic use cases as a primary target. On multi-step autonomous tasks (APEX-Agents), tool coordination (MCP Atlas: 69.2%), and autonomous web research (BrowseComp: 85.9%), the model shows substantial gains over its predecessor. These improvements make it well-suited for AI agent pipelines, complex coding assistants, and long-horizon research workflows.

The model is available in preview via the Gemini API, Google AI Studio, Vertex AI, Gemini Enterprise, Gemini CLI, and Android Studio. It also powers NotebookLM for Google AI Pro and Ultra subscribers.

What This Means

Gemini 3.1 Pro’s ARC-AGI-2 score is significant not just as a number — it represents meaningful progress on a benchmark specifically designed to resist training-data pattern matching. Combined with its doubled agentic performance, the release signals that frontier models are becoming increasingly capable of handling open-ended, multi-step tasks with minimal human intervention.

For researchers and developers at NYU Shanghai, this release also raises the bar for what “capable” means in practical AI tooling. The 1M token context window makes it feasible to feed entire codebases or research document collections into a single query. The new thinking level controls give developers finer-grained cost-performance management — useful when building production-scale AI applications on limited budgets.

Google’s decision to hold the price steady while doubling reasoning performance also intensifies competition across the frontier model landscape, where pricing pressure has become as significant as raw capability.

This post was drafted with AI assistance and reviewed by RITS staff.

Sources

Related stories

AI

Gemini 3: A New Era of Intelligence from Google

On November 18 2025, Gemini 3 was officially introduced by Google DeepMind and Google LLC. (blog.google) This marks what they call “a new era of intelligence” — with Gemini 3 being described as their most advanced model yet. (blog.google) Key points On November 18 2025, Gemini 3 was officially introduced by Google DeepMind and Google […]

cli

Introducing Gemini CLI: Bring AI Power to Your Terminal

On June 25, 2025, Google unveiled Gemini CLI, an open-source AI agent that integrates the capabilities of the Gemini 2.5 Pro reasoning model directly into your command-line environment (theverge.com). This marks a significant shift from web-based AI tools to a native terminal experience, enabling developers to write and debug code, generate content, and perform deep […]

api

Google Unveils Gemini 2.5 Models with Native Text-to-Speech Capabilities

At Google I/O 2025, Google introduced significant updates to its Gemini 2.5 model series, notably integrating native text-to-speech (TTS) capabilities. This advancement positions Gemini as a formidable contender in the AI-generated speech domain, directly challenging offerings like OpenAI’s GPT-4o. Key points At Google I/O 2025, Google introduced significant updates to its Gemini 2.5 model series, […]

Never miss a RITS update

Subscribe to the RSS feed for AI, emerging technology, and digital scholarship news curated for the NYU Shanghai community, or get in touch to suggest a story.