Can Gemini 3.1 Pro Solve Complex Tasks? A Deep Dive into Google’s New AI Model

Overview

Google recently released Gemini 3.1 Pro, a large multimodal model aimed at tackling tasks where simple answers are insufficient. The model emphasizes enhanced core reasoning capabilities and positions itself as a stronger baseline for complex problem solving.

Benchmark Performance

On the ARC‑AGI‑2 benchmark, which evaluates a model’s ability to solve novel logical patterns, Gemini 3.1 Pro achieved a 77.1% validation score, more than doubling the inference performance of the previous Gemini 3 Pro.

Agent Town Experiment

The authors tested the model on an “Agent Town” scenario involving autonomous agents that live, socialize, and share information. While the results were modest and lagged behind Claude’s performance with the same prompts, the experiment highlighted the model’s ability to generate a town map UI and populate it with 25 distinct agents such as a pharmacy clerk, a professor, a music student, and a mayoral candidate.

Town map UI examples: Hobbs Cafe, The Rose & Crown, Oak Hill College, etc.

Agents: John Lin (pharmacy clerk), Mei Lin (professor), Eddy Lin (music student), Sam Moore (mayor candidate), …

For a side‑by‑side comparison with Opus 4.5 and Sonnet 4.6, see the linked article.

Official Demonstrations

Code‑based animation: Gemini 3.1 Pro can generate SVG animations directly from textual prompts. Because the output is pure code, the animations remain crisp at any scale and have a much smaller file size than traditional video.

Complex system synthesis: The model built a real‑time aerospace dashboard that visualizes International Space Station telemetry by stitching together public data streams, showcasing its ability to bridge sophisticated APIs with user‑friendly designs.

Interactive design: A 3D flocking‑bird simulation was created, complete with gesture‑based control and generative music that adapts to the birds’ motion, offering researchers a powerful prototyping tool for multimodal interfaces.

Creative programming: When asked to design a modern portfolio website inspired by Emily Brontë’s Wuthering Heights , Gemini 3.1 Pro inferred the novel’s atmospheric tone, produced a stylish UI concept, and generated functional code for the site.

Model Characteristics

Input context window: 1 million tokens

Output limit: 64 kilo‑tokens

Architecture built on Gemini 3 Pro

Pricing Advantage

The preview tier (gemini‑3.1‑pro‑preview) offers two pricing slabs:

Up to 200 k tokens input → $12.00 per M tokens output

200 k tokens input → $18.00 per M tokens output

Knowledge cutoff is January 2025.

References

https://deepmind.google/models/model-cards/gemini-3-1-pro/

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

https://simonwillison.net/