Google released Gemini 3.1 Pro yesterday, and the benchmark numbers are striking. On ARC-AGI-2, a test of abstract reasoning, the new model scored 77.1%. Its predecessor, Gemini 3 Pro, scored 31.1%. That is more than double the score on this benchmark, not proof that the model’s overall reasoning capability doubled.
Major model releases now arrive with long tables of claimed gains, making it easy for the numbers to blur together. Gemini 3.1 Pro’s pricing is also notable: Google charges $2 per million input tokens for prompts up to 200,000 tokens, matching Gemini 3 Pro’s rate for that tier. Longer prompts cost $4 per million input tokens.
The Benchmark Picture
Google’s own model card shows Gemini 3.1 Pro leading or tying many of the listed evaluations, but not all of them. The standout results include:
- ARC-AGI-2 (abstract reasoning): 77.1%, compared to Anthropic Opus 4.6’s 68.8% and OpenAI GPT-5.2’s 52.9%
- GPQA Diamond (scientific knowledge): 94.3%, outperforming all listed competitors
- SWE-Bench Verified (agentic coding): 80.6%
- BrowseComp (web browsing tasks): 85.9%
- LiveCodeBench Pro: Elo score of 2,887
But benchmark dominance isn’t universal. Anthropic’s Claude Opus 4.6 still leads on Humanity’s Last Exam with tool support at 53.1%, and Gemini 3 Pro actually beats its successor on MMMU Pro (81.0% vs 80.5%) - a reminder that newer doesn’t always mean better across every dimension.
What’s Actually New
The model accepts up to 1 million tokens of input context and can produce up to 64,000 tokens of output. It’s natively multimodal, processing text, images, audio, video, and code repositories in a single context.
One genuinely new capability: Gemini 3.1 Pro can generate animated SVGs directly from text descriptions. These vector graphics remain sharp at any size with minimal file sizes - useful for developers building web interfaces.
The model also introduces a three-tier “thinking level” system. Users can choose low, medium, or high computational effort, giving more control over the tradeoff between speed and reasoning depth. The previous generation only offered low and high options.
The Price War Continues
At $2 per million input tokens for prompts up to 200,000 tokens, and $12 per million output tokens in that tier, Gemini 3.1 Pro costs less than Anthropic’s Opus API. Above 200,000 input tokens, Google’s rates rise to $4 per million input tokens and $18 per million output tokens. The benchmark improvements put pressure on competitors to match the price-to-performance ratio or justify higher prices with other capabilities.
Google’s broader business may give it more room to sustain low API prices than AI-only companies have. Whether that leads to durable competition or eventual market consolidation remains an open question.
What This Means for Users
For developers and enterprises already using Google’s AI stack, the model is available in preview through the Gemini API, Google AI Studio, and Vertex AI at the same short-prompt input price as Gemini 3 Pro. It is also rolling out in the Gemini app. NotebookLM access requires a Pro or Ultra subscription.
For everyone else, the practical question is whether benchmark improvements translate to real-world utility. A model that scores 77% instead of 68% on abstract reasoning tests might matter for some applications and not at all for others.
Privacy Considerations Remain
Google’s enterprise offerings - Vertex AI and Gemini for Workspace - do not use customer data to train models.
The consumer Gemini app has different rules. With Keep Activity on, chats are auto-deleted after 18 months by default, though users can change that period to three or 36 months. Google says a subset of chats is reviewed by humans, and reviewed chats can be retained for up to three years even after a user deletes their activity.
Privacy depends on the user’s settings, not simply whether the account is free or paid. Google says future chats are not used to improve its AI models when Keep Activity is off and the user does not submit feedback, though those chats are retained with the account for 72 hours. Users should still avoid entering confidential information they would not want a reviewer to see.
The Bigger Picture
Google’s benchmark table compares Gemini 3.1 Pro with Claude Opus 4.6, Claude Sonnet 4.6, GPT-5.2, GPT-5.3-Codex, and Gemini 3 Pro. The results are mixed rather than a clean sweep: Gemini leads some reasoning, science, coding, and browsing tests while trailing on several agentic and software-engineering evaluations.
Google’s strategy here is to compete on price while matching or exceeding rivals on selected capabilities. The company is betting that lower API prices will drive adoption across the products where Gemini 3.1 Pro is available.
For users, the pace of improvement creates a different challenge: deciding when capabilities are good enough to build on, versus waiting for the next model that might be better. Gemini 3.1 Pro’s ARC-AGI-2 score shows that at least one reasoning benchmark still has room for large jumps. Whether those gains transfer to a user’s workload requires testing.