Google Officially Releases Gemini 3.1 Pro: ARC-AGI-2 and Product Positioning Baseline

One-sentence takeaway

Google positions Gemini 3.1 Pro as a core model for complex problems, multimodal reasoning, and agentic workflows, and reports an ARC-AGI-2 verified score of 77.1%; however, the release page provides only limited benchmark context.

Test environment

• Version: Gemini 3.1 Pro, released as a preview on 2026-02-19.

• Access points: Gemini API/AI Studio, Gemini CLI, Google Antigravity, Android Studio, Vertex AI, Gemini Enterprise, Gemini app, and NotebookLM.

• Benchmark: ARC-AGI-2, which the official description says tests a model's ability to solve entirely new logical patterns.

• Configuration: The release page calls this a verified score but does not disclose the complete prompts, sample split, sampling parameters, or tool traces in the article.

Inputs/configuration

The official release page does not provide the complete ARC-AGI-2 inputs or execution harness. At the product-example level, Google presents the model for complex problems, multimodal interpretation, data synthesis, and creative projects; the developer preview path is gemini-3.1-pro-preview.

Results

• ARC-AGI-2 verified: 77.1%.

• Google says this score is more than twice that of Gemini 3 Pro; the release page does not provide Gemini 3 Pro's complete configuration information in the same article.

• Product positioning: advanced reasoning across complex problems and modalities; at launch, Google explicitly said it would continue validating ambitious agentic workflows before gradually moving toward general availability.

Conclusion

The ARC-AGI-2 result supports treating Gemini 3.1 Pro as a candidate with strong abstract reasoning; if the user's task involves coding agents, SQL, or multi-tool orchestration, it must be evaluated separately at the task level, and cannot be inferred from a single ARC score.

Limitations

• This is an official self-report; the release article does not provide complete raw samples, prompts, random seeds, costs, failure types, or confidence intervals.

• “Verified” indicates that a verification process took place, but it is still not equivalent to an independent third-party reproduction.

• The preview model's API, pricing, rate limits, and behavior may change; the launch-day state should not be treated as a permanent specification.

• The article explicitly presents agentic workflows as an area for continued validation, so it cannot be claimed that the model's Agent capabilities are already comprehensively stable.

Reproduction steps

1. Use a fixed gemini-3.1-pro-preview snapshot and the official permitted ARC-AGI-2 evaluation protocol.

2. Record the thinking level, temperature, tools, input version, outputs, and time spent per question; do not mix Gemini app results with API results.

3. Build separate coding, SQL, and multi-tool task sets, and report success rate, tool-call accuracy, latency, and cost as separate metrics.

4. Use the official 77.1% as the release baseline, clearly indicating whether your own harness is comparable.