Gemini 3.1 Pro claims top-tier reasoning gains

Google has released Gemini 3.1 Pro, a new version of its “Pro” model line that it says upgrades the core reasoning capabilities behind recent Gemini 3 advances. ^[1]

The company is rolling it out starting February 19, 2026 across consumer products and developer and enterprise platforms. 

Context and background

Gemini 3.1 Pro arrives about a week after Google DeepMind’s update to Gemini 3 Deep Think, which the company positioned as a research-focused system for science and engineering workflows. ^[2]

Google describes 3.1 Pro as the “core intelligence” behind those improvements, now packaged for broader day-to-day use. 

Key details

Google says Gemini 3.1 Pro is designed for tasks where a short, direct answer is not enough, including multi-step reasoning, synthesis, and complex problem solving. 

Availability (starting today):

• Developers (preview): via the Gemini API in Google AI Studio, plus access through Gemini CLI, Android Studio, and Google’s agentic development tooling (including “Antigravity,” per Google’s announcement). 

• Enterprise: via Vertex AI and Gemini Enterprise. 

• Consumers: via the Gemini app and NotebookLM, with higher usage limits tied to Google’s paid tiers. 

Benchmark claims

Google highlights performance on ARC-AGI-2, reporting a verified score of 77.1% for Gemini 3.1 Pro and describing it as more than double Gemini 3 Pro’s reasoning performance on that benchmark. 

Google also publishes a broader benchmark table for Gemini 3.1 Pro across reasoning, coding, agentic tool use, multilingual evaluation, and long-context testing. 

Google also showcased the new model’s SVG generation capabilities (from the official Google blog):

I used the new model to create an animated hummingbird SVG and it was surprisingly good:

Why this matters

For most people, the practical question is not whether a model can answer trivia. It is whether it can reliably complete multi-step work without losing the thread.

Google is positioning Gemini 3.1 Pro as a stronger default model for those longer tasks, while simultaneously pushing it into products where users expect it to be dependable, including NotebookLM for research workflows and Vertex AI for production deployments. 

The benchmark emphasis also signals where Google wants to compete: not only on general reasoning, but on agent-style evaluation categories such as tool use and coding. The company’s publication of an evaluation methodology document makes it easier to compare results, but it does not eliminate the need to examine how each benchmark was run and whether conditions match real-world use. 

Keeping up with other SOTA model releases

Over the past few weeks, the frontier model market has moved quickly, with major labs shipping new flagships and specialized variants in short succession. OpenAI introduced GPT-5.3-Codex and later GPT-5.3-Codex-Spark, positioning them around agentic coding and faster interactive development workflows. 

Anthropic has also updated its top tier models, releasing Claude Opus 4.6 (and a fast mode for it) and then, in the same month, releasing Claude Sonnet 4.6; with both announcements emphasizing stronger reasoning and coding performance, plus support for long-context work. 

Google’s release of Gemini 3.1 Pro fits into the same pattern. The company is framing 3.1 Pro as a step up in “core reasoning,” and is rolling it out broadly across consumer, developer, and enterprise surfaces rather than limiting it to a single product. It also comes just a day after Google began rolling out Lyria 3, a built-in tool for generating short music tracks, in the Gemini app.