Gemini 3.1 Pro: Benchmarks, Animated SVGs, and CAD Demos

It is Gemini day: Google has released Gemini 3.1 Pro. I want to start with what I jokingly call Pelican Bench: an animated SVG of a pelican riding a bicycle. The newer example looks more detailed and fluid than the Gemini 3 Pro version. Google also shows a frog on a penny-farthing, a giraffe in a tiny car, and an ostrich on roller skates. My AGI remark is a joke; these are illustrative visual examples, not a measure of general intelligence.

Google also said the Deep Think upgrade announced the previous week used the newer Gemini 3.1 core, and is bringing 3.1 Pro to consumer and developer products. I am eager to try it in OpenClaw. The launch figures that follow are Google’s reported comparisons from February 2026, so their settings and the date matter.

On Humanity’s Last Exam, Google reports 44.4% for Gemini 3.1 Pro without tools and 51.4% with code and search subject to a blocklist. Both improve on the displayed Gemini 3 Pro figures. The table places it ahead of Opus 4.6 without tools but a little behind it with tools; the two numbers within each column are not directly interchangeable because the setups differ.

ARC-AGI-2 tests abstract visual reasoning with examples of a pattern and its solution, then asks the model to solve a new instance. Google reports an ARC Prize verified 77.1% for Gemini 3.1 Pro, compared with 31.1% for Gemini 3 Pro and 68.8% for Opus 4.6 in the table. It is a striking result on that test, though a benchmark score does not establish that a system has achieved AGI.

For GPQA Diamond, the table reports 94.3% with no tools. On SWE-bench Verified, Gemini 3.1 Pro gets 80.6% in a single-attempt setup, near the listed 80.8% for Opus 4.6 and ahead of the older Gemini 3 Pro result. These are task-specific comparisons under the configurations shown, not guarantees of performance on every scientific or coding task.

The other interesting model in the table is Anthropic’s Sonnet 4.6. In my recent experience it has been a workhorse for knowledge work, with a personality I like, strong tool calling and a million-token context window. At the time of this video I find it expensive, although less so than Opus, which is why I am not using it as my primary OpenClaw model. That is my personal assessment and the pricing context may change.

Google also lists strong results on LiveCodeBench and SciCode, and 99.3% on the telecom portion of τ2-bench, which measures agentic tool use. But benchmarks only tell part of the story. What matters in practice is how a model behaves when you use it. I briefly switched to Gemini 3 Pro in OpenClaw after my Anthropic OAuth access was revoked, liked it, and want to see how 3.1 performs.

Google describes 3.1 Pro as intended for problems requiring more than a simple answer. Its chameleon SVG comparison shows a more detailed design and eyes intended to track the cursor. I am not sure why the launch puts such emphasis on SVGs, but the visual difference in the examples is striking. These are selected demonstrations and individual results can vary.

Another example is a simulation of starling murmuration: a flock moving together in shifting patterns. Google presents the interactive simulation as something built with the model. It illustrates a creative coding application, though this demonstration alone does not establish that it faithfully models real bird behavior.

Google DeepMind chief scientist Jeff Dean also shared an urban-planning application created with Gemini 3.1 Pro. In the demonstration, it designs a city around a landscape and offers settings for paths, collisions, water and land features. It builds city blocks and produces an image suggesting what the result could look like. I find the interface impressive, even if a demo is not evidence that its plans are suitable for a real city.

The final example catches my attention because I have been getting into 3D printing. A prompt asks for ManifoldCAD code to recreate an object from a technical drawing, and the demonstration shows the generated code and a rendered three-dimensional part. I imagine using this approach for many designs; the example does not verify every measurement, however, so a person would still need to inspect dimensions, geometry and manufacturing suitability before printing or using a part.