Short answer
Gemini 3.1 Pro, released by Google in preview on February 19, 2026, is an upgrade of Gemini 3 Pro focused on core reasoning, scoring a verified 77.1% on the ARC-AGI-2 puzzle test, more than double its predecessor. It is rolling out in the Gemini app and NotebookLM (with higher limits for AI Pro and Ultra subscribers) and to developers through the Gemini API, Vertex AI and Google's coding tools.
Follow-up question How can you try Gemini 3.1 Pro today?
Yesterday, Thursday, February 19, 2026, Google released Gemini 3.1 Pro. It comes three months after Gemini 3 Pro and a week after Google updated Gemini 3 Deep Think, its slower, research-grade reasoning mode. Google describes 3.1 Pro as the upgraded "core intelligence" behind those Deep Think results, now brought to everyday products.
It is officially a preview: Google says it wants to validate the improvements, especially for agent-style work, before making the model generally available "soon." In practice, preview means it is usable now but may still change.
What Google released
Gemini 3.1 Pro is rolling out on three fronts at once:
- Consumers: the Gemini app, with higher usage limits for Google AI Pro and Ultra subscribers, and NotebookLM, Google's research notebook, for Pro and Ultra users only.
- Developers: the Gemini API in Google AI Studio, the Gemini CLI command-line tool, Google Antigravity (Google's agent-based coding platform) and Android Studio.
- Businesses: Vertex AI and Gemini Enterprise.
According to its model card, the published technical summary, Gemini 3.1 Pro is built on Gemini 3 Pro and is natively multimodal: it takes text, images, audio, video and whole code repositories as input. It accepts up to 1 million tokens and writes up to 64,000 tokens in a reply. A token is a small piece of text, often part of a word.
How it compares
The headline number is ARC-AGI-2, a test of abstract puzzles that each follow a new pattern the model has never seen. It is designed so that memorized knowledge doesn't help; the model must work out the rule. Here are some scores from Google's model card, each with the thinking setting Google lists for it (High for Gemini, Max for Claude, xhigh for GPT-5.2):
| Test | What it checks | Gemini 3.1 Pro | Gemini 3 Pro | Claude Opus 4.6 | GPT-5.2 |
|---|---|---|---|---|---|
| ARC-AGI-2 | New logic puzzles | 77.1% | 31.1% | 68.8% | 52.9% |
| Humanity's Last Exam (no tools) | Expert questions | 44.4% | 37.5% | 40.0% | 34.5% |
| GPQA Diamond | Graduate-level science | 94.3% | 91.9% | 91.3% | 92.4% |
| SWE-Bench Verified | Fixing real code bugs | 80.6% | 76.2% | 80.8% | 80.0% |
| BrowseComp | Finding hard facts online | 85.9% | 59.2% | 84.0% | 65.8% |
The ARC-AGI-2 jump, from 31.1% to 77.1%, is the reason Google says reasoning has more than doubled. On coding, the picture is closer: Opus 4.6 edges ahead on SWE-Bench Verified. The table also shows a weak spot: on GDPval-AA, which rates professional tasks, Gemini 3.1 Pro trails both Claude models. As always, these are a company's own numbers, chosen and run by Google.
What the reasoning looks like in practice
Google's examples are mostly about turning a request into working code:
- Animated graphics from text. It writes SVG animations, pictures described in code rather than pixels, so they stay sharp at any size and the files stay tiny.
- Connecting live data. It built a dashboard that pulls a public telemetry feed to show the International Space Station's orbit in real time.
- Interactive design. It coded a 3D flock of starlings that users can steer with hand gestures, with music that changes as the birds move.
- Interpreting a theme. Asked for a portfolio website for the main character of Wuthering Heights, it reasoned about the novel's mood to design the page.
The common thread is a request that needs several steps of thought before any output: understanding an API, planning a layout, then writing correct code.
Safety
Google tested 3.1 Pro under its Frontier Safety Framework, which checks for serious risks such as help with chemical or biological weapons, cyberattacks, manipulation and AI research automation. The model card reports that it stays below the "critical capability levels" in all areas. In cybersecurity it has reached an early-warning alert threshold, as Gemini 3 Pro did, while remaining below the critical level, so Google keeps extra protections in place. Google also says it outperforms Gemini 3 Pro overall on its automated safety and tone tests while keeping unjustified refusals low.
Try it: In the Gemini app, ask for something that needs both reasoning and code, such as "Create an animated SVG of a sunrise over three hills, with the sky changing color over ten seconds. Give me the full code." Paste the result into a file ending in .svg and open it in your browser.
How can you try Gemini 3.1 Pro today?
If you use the Gemini app, look for 3.1 Pro in the model selection as it rolls out. Paying Google AI Pro and Ultra subscribers get higher limits, and they can also use it inside NotebookLM with their own documents.
Developers can test it immediately in Google AI Studio by choosing the 3.1 Pro preview, or call it through the Gemini API, Vertex AI or the Gemini CLI. Because it is a preview, keep your production apps on a stable model for now and run 3.1 Pro side by side on real tasks.
A good first test is a reasoning problem from your own work:
Here is last quarter's sales table. Find the two most surprising
patterns, explain how you found them step by step, and suggest
one chart that would show each pattern clearly.Check the steps, not just the answer. Strong reasoning models are most useful when you can follow, and verify, how they reached their conclusion.
Sources
- Gemini 3.1 Pro: A smarter model for your most complex tasks — Google, February 19, 2026
- Gemini 3.1 Pro model card — Google DeepMind, February 19, 2026