Gemini 3.1 Pro: How Google Boosted Reasoning Scores and What It Means for Developers

Google announced the Gemini 3.1 Pro model, emphasizing a major leap in inference capability compared with the previous Gemini 3 Pro.

Benchmark Improvements

ARC‑AGI‑2 (abstract reasoning) rose from 31.1% to 77.1% .

GPQA Diamond (scientific knowledge reasoning) increased from 91.9% to 94.3% .

Terminal‑Bench 2.0 (terminal programming) improved from 56.9% to 68.5% .

SWE‑Bench Verified (code repair) went from 76.2% to 80.6% .

Preview Availability

The model is currently in preview and has been pushed to the Gemini API (AI Studio), Gemini CLI, Antigravity, Vertex AI, Gemini App, and NotebookLM.

Official Demo Highlights

SVG animation with finer detail than Gemini 3 Pro.

Real‑time data dashboard visualizing International Space Station telemetry, from API integration to page rendering in a single flow.

3D interactive simulation using a "3D hoid" algorithm to create a hand‑controlled flock‑of‑birds experience.

Literary‑to‑website conversion: the model generated a high‑quality personal site based on the novel Wuthering Heights .

Community Demos by Jeff Dean

Google’s chief scientist Jeff Dean showcased several cases:

SVG animation.

City‑planning simulation to design an entirely new metropolis.

Heat‑conduction analysis using Gemini 3.1 Pro’s Deep Think capability, performed without external tools. The workflow consists of three steps: (1) generate a CAD model from technical drawings, (2) run thermal analysis based on CAD and material parameters, and (3) render temperature‑field visualizations for different heating times.

Pricing

Standard context (≤200 K tokens): input $2 , output $12 .

Extended context (>200 K tokens): input $4 , output $18 .

Access Channels

Developers: preview available through Gemini API (Google AI Studio, Gemini CLI, Google Antigravity, Android Studio) for building intelligent agents and applications.

Enterprises: reachable via Vertex AI and Gemini Enterprise for large‑scale workloads.

General users: offered through Gemini App and NotebookLM, currently limited to Google AI Pro and Ultra subscription tiers with higher usage quotas.

References

Official blog:

https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-1-pro/

Benchmark methodology:

https://deepmind.google/models/evals-methodology/gemini-3-1-pro

Gemini API documentation:

https://ai.google.dev/gemini-api/docs