Google is kicking off 2026 with exactly the kind of AI flex you’d expect: a fresh flagship model aimed squarely at “hard mode” problems, not just chatty summaries and homework help. The new Gemini 3.1 Pro is Google’s latest Pro‑tier model in the Gemini 3 family, pitched as the smarter baseline for complex reasoning and agent-style workflows across its consumer, developer, and enterprise stack.

On paper, the headline stat is eye‑catching. Gemini 3.1 Pro posts a verified 77.1% score on ARC‑AGI‑2, a notoriously tough benchmark designed to test whether a model can solve entirely new abstract logic puzzles rather than regurgitate patterns it has seen before. Google says that’s more than double the reasoning performance of Gemini 3 Pro, which until recently was its top general-purpose reasoning model. Put simply, this is the model Google wants you to reach for when a “quick answer” is not enough and you actually need the AI to think.

The launch comes just a week after Gemini 3 Deep Think, an even more compute‑hungry variant that pushed ARC‑AGI‑2 up to 84.6% and sparked fresh “is this AGI yet?” debates among AI watchers. Deep Think is framed as the research and science workhorse, focusing on test‑time compute — letting the model “think longer” before answering — while 3.1 Pro is the more practical core intelligence that trickles into everyday products and APIs. In other words, Deep Think is the lab car breaking lap records; Gemini 3.1 Pro is the production model you can actually drive.

Google is not keeping this one locked in a research paper. From day one, Gemini 3.1 Pro is rolling out broadly in preview. Developers can hit it via the Gemini API in Google AI Studio, the Gemini CLI, Android Studio, and Google’s new agentic development environment, Antigravity. Enterprises get it through Vertex AI and Gemini Enterprise, tying into the broader Google Cloud stack. And on the consumer side, the model is arriving in the Gemini app and in NotebookLM, though with a catch: the higher limits and access land first for paying Google AI Pro and Ultra subscribers. The strategy is clear: this is the “default brain” Google wants permeating everything from hobby projects to corporate workflows — but the best access tiers are increasingly paywalled.

Under the hood, Gemini 3.1 Pro is an evolution of Gemini 3 rather than an entirely new species. Gemini 3 already introduced a strong reasoning-first architecture with a huge context window (up to around 1 million tokens in Gemini 3 Pro), multimodal support across text, code, images, audio, video, and long documents, and a “thinking level” control so developers can trade off latency and cost against deeper internal reasoning. 3.1 Pro leans into that template but pushes the reasoning bar higher, especially on abstract pattern recognition and complex multi‑step tasks, which ARC‑AGI‑2 is designed to probe.

Google’s own examples focus less on raw benchmarks and more on what this looks like in practice. One scenario highlights 3.1 Pro generating website‑ready animated SVGs directly from a text prompt. Because those animations are produced as pure code rather than pixels, they scale cleanly at any resolution and weigh far less than traditional video, which matters for performance‑sensitive sites or UI prototypes. Another example shows the model wiring up a live aerospace dashboard by configuring a public telemetry stream to visualize the International Space Station’s orbit, effectively bridging between low‑level APIs and a user‑friendly front‑end without hand‑holding. These are carefully chosen demos, but they hint at the direction: from “answer my question” to “build and wire the system that answers my question.”​

That’s where Antigravity enters the picture. Google’s new agent‑first development platform is built specifically to sit on top of models like Gemini 3 and 3.1 Pro, giving them direct access to the code editor, terminal, and browser so they can plan and execute multi‑step software tasks with relatively little human intervention. Instead of a chat window bolted onto your IDE, Antigravity frames agents as active collaborators that can spin up in parallel workspaces, do background research, modify code, run tests, and validate their own changes in the browser. Plug a more capable reasoning model into that environment and you get a glimpse of Google’s bet: AI that not only explains complex systems but autonomously glues them together.

From a product standpoint, Gemini 3.1 Pro is also a move in Google’s ongoing platform chess match. Gemini 3 Pro already underpins a growing list of integrations, from AI Studio and Vertex AI to third‑party tools like Cursor, GitHub integrations, JetBrains, Replit, and more. By slotting 3.1 Pro in as the new high‑intelligence default, Google effectively raises the baseline for what “Pro” experiences and cloud APIs can do without forcing developers to redesign their entire stack. In Vertex AI, this means customers who were already using Gemini 3 Pro for long‑context retrieval, document analysis, or tool‑calling agents can begin experimenting with 3.1 Pro to see if it unlocks more reliable reasoning and complex workflows.

On the consumer side, tying 3.1 Pro access to subscription plans is another sign of the emerging “tiered intelligence” model across the industry. Google’s AI Pro and Ultra offerings are marketed around giving users the highest access to Gemini 3‑class models, plus perks like Deep Search, more powerful agent features, and interactive tools for complex questions. With 3.1 Pro rolling out first to those paying tiers inside the Gemini app and NotebookLM, Google is effectively telling power users — from students working on serious research to professionals building knowledge workflows — that the smartest model is part of a premium bundle.

The timing of the release is also notable in the broader AI arms race. ARC‑AGI‑2 has quickly become one of the headline benchmarks for “frontier” models because it tries to measure something closer to general problem‑solving intelligence rather than narrow exam performance. Models like Gemini 3 Deep Think, which reaches into the mid‑80s on ARC‑AGI‑2, and rivals such as specialized systems highlighted by the ARC Prize community have pushed those numbers aggressively upward in the last year. But improvements on more traditional benchmarks, such as coding leaderboards or language understanding, have been more incremental, which has sparked a debate about how well ARC‑AGI‑style scores translate into real‑world robustness. With 3.1 Pro, Google is effectively using ARC‑AGI‑2 as a marketing signal while trying to ground the story in very practical, product‑driven capabilities.

At the same time, Google is careful to label Gemini 3.1 Pro as “preview” across its developer and enterprise channels. The company says it wants to validate the updates and continue iterating on areas like ambitious agentic workflows before declaring the model generally available. That’s an implicit acknowledgment that this generation of systems is being refined in public, with production environments and early adopters serving as a kind of extended testbed. It also gives Google room to tune pricing, performance, and safety systems as more real‑world usage data comes in, especially for high‑stakes or automated use cases.

For developers, the immediate takeaway is straightforward: there’s a new Pro‑tier Gemini model to experiment with, accessible through the same tools they’re already using, but with significantly stronger performance on difficult reasoning tasks and the potential to power more capable agents. For enterprises, it is one more step in Google’s pitch that its cloud is not just a place to host models but a full stack for building AI‑native applications, from data to orchestration to user interface. And for everyday users, most of whom will meet 3.1 Pro through the Gemini app or NotebookLM, the change may be less about raw benchmarks and more about subtle shifts: better explanations of complex topics, smarter synthesis of sprawling information, and AI‑powered tools that feel a bit more like collaborators than autocomplete.

None of this settles the big questions around AI — what counts as “general intelligence,” how to measure it, or how far we are from something qualitatively different. But Gemini 3.1 Pro is a clear marker of where Google wants to go: models that don’t just answer questions, but reason through them, manipulate tools and code on your behalf, and quietly power a new layer of agentic software behind almost every interface you touch.