Gemini 3.1 Pro leads with browser-native code generation, not just chat answers

What Google actually shipped, and where

On February 19, 2026, Google released Gemini 3.1 Pro in preview, calling it the "upgraded core intelligence" behind last week's Gemini 3 Deep Think update. It is rolling out across the Gemini API in AI Studio, Gemini CLI, the agentic development platform Antigravity, and Android Studio for developers; Vertex AI and Gemini Enterprise for enterprises; and the Gemini app and NotebookLM for consumers.

The headline number is a single benchmark: 77.1% on ARC-AGI-2, which Google describes as testing a model's ability to solve entirely new logic patterns, and which it says is more than double the reasoning performance of Gemini 3 Pro. Notably, that is the only quantified claim in the post. Everything else Google chose to show is rendered in a browser.

Consumer access is gated: higher limits in the Gemini app go to Google AI Pro and Ultra plan users, and NotebookLM access is exclusive to those same tiers. Google frames the preview as a validation step "before we make it generally available soon," with agentic workflows named as the area still being pushed.

Four demos that are all really frontend demos

The examples Google leads with are unusually concrete about the frontend. First, code-based animation: the model generates website-ready animated SVGs directly from a text prompt. Google's argument is specific — because these are pure code rather than pixels, they stay crisp at any scale and carry small file sizes compared to video. That is a claim aimed squarely at people who care about asset weight and rendering, not at people impressed by a picture.

Second, a live aerospace dashboard that configures a public telemetry stream to visualize the International Space Station's orbit. Google's framing is that the model "bridges the gap between complex APIs and user-friendly design" — meaning it wired a real data source into a working interface, not a static mockup.

Third, a 3D starling murmuration that responds to hand-tracking and plays a generative score that shifts with the birds' movement. Google explicitly pitches this at "researchers and designers" who want to prototype sensory-rich interfaces. Fourth, a portfolio site for Emily Brontë's "Wuthering Heights" where the model reasoned through the novel's tone to produce a contemporary layout rather than a summary.

3.1 Pro is designed for tasks where a simple answer isn't enough, taking advanced reasoning and making it useful for your hardest challenges.Montana Labs

Reasoning gains presented as generated interface, not generated text

What's worth noticing is the throughline Google chose. A model vendor could demonstrate a doubled ARC-AGI-2 score with math proofs, scientific synthesis, or long-context analysis. Instead, three of the four showcased outputs are things you view in a browser, and the fourth — the dashboard — is a browser view backed by a live API call. The pitch is that better reasoning shows up as better structured code: SVGs that stay small, dashboards that correctly parse a telemetry stream, WebGL scenes that stitch together interaction, audio, and rendering.

For teams building tools that turn prompts into shippable UI, this is the useful signal. The failure modes in generated frontends are rarely "the model didn't know CSS." They are broken data bindings, oversized or malformed assets, and interactions that look right in a screenshot but don't run. Google is asserting, through its example selection, that 3.1 Pro's reasoning step reduces exactly those integration failures — configuring a real stream, keeping file sizes down, coordinating hand-tracking with generative audio.

The preview label is the part to test against

The concrete implication for anyone integrating 3.1 Pro: this is a preview whose most impressive claims are frontend code that runs, so validate against that same bar. Google says it is releasing the model to validate updates and advance agentic workflows before general availability. The demos set an expectation — animated SVGs at small file sizes, a working telemetry visualization, a functioning 3D interaction — that a curated example gallery cannot confirm at your scale.

If you build on 3.1 Pro through the API, Antigravity, or Android Studio during preview, the honest test is whether generated SVGs actually stay crisp and lightweight in production, whether API wiring holds up beyond a public ISS feed, and whether interactive scenes degrade gracefully across devices. Google gave four falsifiable frontend claims. Treat them as a checklist, not a portfolio.