Gemini 3.1 Pro Just Dropped. And It's Gunning for the Top Spot

If you blinked yesterday, you might have missed it. Google quietly (and then not so quietly) released Gemini 3.1 Pro on February 19, and the numbers are turning heads. This isn't a minor patch or a marketing reskin. The benchmark jump from Gemini 3 Pro to 3.1 Pro is the kind of leap that makes you want to re-evaluate what's sitting in your tech stack.

Let's break down what actually changed, why it matters for developers, and what the arrival of Google's free Antigravity IDE means for your AI-assisted workflow.

The Benchmark Jump That Raised Eyebrows

The headline number is the ARC-AGI-2 score, a test designed to measure genuine problem-solving and reasoning ability rather than pattern matching. Gemini 3 Pro scored 31.1%. Gemini 3.1 Pro scores 77.1%.

That's not an incremental improvement. That's more than double, in a single point release. For context, Claude Opus 4.6 sits at 80.8% on the same test. Gemini 3.1 Pro is now breathing down its neck.

On SWE-Bench Verified, the benchmark that tests how well a model can resolve real GitHub issues in actual software projects, Gemini 3.1 Pro scores 80.6%. Opus 4.6 scores 80.8%. We're now talking about a statistical dead heat between the two best coding models available.

The practical upshot? You now have a genuine frontier-class alternative to Claude for complex agentic coding work. And critically, Gemini 3.1 Pro is priced at $2/million input tokens and $12/million output tokens, roughly half the cost of Opus 4.6. If you're running agentic workflows at any kind of scale, that cost difference adds up fast.

The Output Window Change Nobody's Talking About Enough

Everyone obsesses over context in, specifically how much the model can read. But in agentic workflows, the output limit is often the real bottleneck. Gemini 3.1 Pro ships with a 65,000 token output limit. That's a significant jump, and here's why it matters in practice.

Imagine asking your agent to generate a complete multi-module feature (controllers, services, tests, documentation) in a single pass. Previously you'd hit an output ceiling mid-generation and need to break the job into multiple sequential calls, each requiring careful hand-off of state. With 65K output tokens available, you can now generate a full, cohesive implementation in one shot. That reduces orchestration complexity, minimises context loss between calls, and makes your agentic pipelines meaningfully simpler to build.

For anyone building long-form generation tasks like API scaffolding, comprehensive technical specs, or full test suites, this is the change that should get your attention.

The 1M Token Context Window: Feed It Your Whole Codebase

The 1M token context window isn't new to the Gemini family, but combined with the reasoning improvements in 3.1, it becomes more useful. You can now feed the model an entire medium-sized codebase and have it reason across cross-file dependencies without losing the thread.

Google has also bumped the file upload limit from 20MB to 100MB via the API, and added direct YouTube URL support as a media source, so the model can now "watch" a video via URL rather than requiring you to upload it manually. For developers building anything that involves video content or large document processing, these are meaningful quality-of-life improvements.

One API Breaking Change to Watch

If you're already building on the Gemini API, there's a small but important change to know about. In the Interactions API v1beta, the field total_reasoning_tokens has been renamed to total_thought_tokens.

It's a minor rename, but it signals something architecturally significant. Google is doubling down on "thought signatures", which are encrypted representations of the model's internal reasoning that get passed back to the model in multi-turn agentic workflows. This is how the model maintains context and coherence across a long sequence of agent actions. If you're building multi-turn agentic flows on the Gemini API, update your code to use the new field name, and it's worth reading up on how thought signatures work. They'll become increasingly central to how you architect Gemini-based agents.

Enter Antigravity: Google's Free IDE Play

Now let's talk about the other thing Google released alongside 3.1 Pro: Antigravity, their VS Code-based IDE. It's free. It uses parallel agent orchestration, meaning it can run multiple coding tasks simultaneously rather than forcing them to queue up sequentially. Fix a bug, refactor a module, and write tests, all at once.

Cursor currently charges $20/month. Windsurf recently cut to $15/month to stay competitive. Google walked in at zero.

For individual developers, that's an obvious conversation. For engineering teams managing AI tooling across 50, 100, or 200 developers, the economics shift dramatically. At $20/month per seat, a 200-person team is spending $48K/year on AI IDE subscriptions alone. Antigravity flips that to zero during the public preview period.

The honest caveat is that it's still in preview, and preview-stage tools carry preview-stage rough edges. Early users are already reporting high-demand errors and slow response times, the classic signs of a popular launch stretching infrastructure. But the architecture is sound, the underlying model is now clearly tier-one, and Google has the scale to back it.

If you're on Cursor or Windsurf and happy with the workflow, there's no urgent reason to switch today. But it's worth running a side-by-side evaluation now rather than waiting until the pricing conversation becomes unavoidable. Antigravity is also available as an API integration point for developers building on the Gemini stack, so if you're already using Gemini in your projects, the tooling integration story is worth exploring.

Gemini 3.1 Pro is available right now via the Gemini API in Google AI Studio, Gemini CLI, Antigravity, Android Studio, Vertex AI, Gemini Enterprise, and via GitHub Copilot and Visual Studio Code.

The Bigger Picture

What Gemini 3.1 Pro signals isn't just that Google has a better model. It's that the gap between frontier models is closing to the point where price, ecosystem fit, and workflow integration matter more than raw capability. When two models are within 0.2% of each other on the most demanding coding benchmark, the deciding factor is how well they plug into your existing stack, how much they cost at your usage volume, and how the tooling around them supports your team.

That's a healthier market than the one we had twelve months ago, where one or two models were clearly ahead of the pack. Developers now have genuine choices at the frontier. Worth taking the time to evaluate what that means for your stack.

Gemini 3.1 Pro is live in preview now. Antigravity is available as a free public preview. The total_reasoning_tokens to total_thought_tokens API rename is effective immediately for the v1beta Interactions API.