In late February, Google DeepMind released Gemini 3.1 Pro. Since then, one number has been widely shared in the developer community: of the 18 major benchmarks currently tracked, 3.1 Pro took first place in 12.
More noteworthy is ARC-AGI-2 — a test specifically designed to evaluate a model's ability to solve "never-before-seen novel logic problems," preventing models from memorizing training data to pass. 3.1 Pro scored 77.1% on this benchmark.
What a Doubling of Reasoning Capabilities Actually Means
Google officially states that 3.1 Pro's reasoning ability is more than twice that of Gemini 3 Pro.
That statement sounds vague, but it is convincing when combined with the ARC-AGI-2 result. ARC-AGI presents entirely new problems each time, requiring the model to infer patterns from a few examples and apply them — there is no way to rely on memorized answers, only genuine reasoning.
A 2x improvement roughly means: given a complex, multi-step problem, the model can decompose, deduce, and verify on its own without needing step-by-step guidance in the prompt. A new MEDIUM Thinking Level has been added (previously only on/off), allowing control over reasoning depth — not every task requires the deepest reasoning, so it can be adjusted as needed.
Technical Specifications
Context and Output:
- Context window: 1 million tokens
- Maximum output: 65K tokens
Supported Input Formats:
- Text and code
- Images (up to 3,000 per request, max 7MB each)
- Audio (up to 8.4 hours)
- Video (approximately 45 minutes with audio, approximately 1 hour without audio)
- PDF (up to 3,000 pages per request, max 50MB per file)
What can 1 million tokens hold? Roughly the entire codebase of a large software project, or 900 pages of PDF, or 8.4 hours of podcast audio. Dropping an entire project in and asking questions directly is now genuinely feasible.
Practical Features
Beyond reasoning, 3.1 Pro adds several practical capabilities:
- Grounding with Google Search: can perform real-time searches during responses, reducing issues from knowledge cutoff dates
- Code execution: the model can run code to verify results, not just generate code
- Finance and Spreadsheet optimizations: these two agentic scenarios have been specifically tuned
- RAG Engine integration: easier connection to enterprise knowledge bases
It also supports batch prediction, and costs are significantly lower for long-context tasks when using caching.
Pricing
| Item | Price |
|---|---|
| Input | $2.00/M tokens |
| Output | $12.00/M tokens |
| Reasoning mode input | $12.00/M tokens |
| Cache read | $0.20/M tokens |
| Cache write | $0.38/M tokens |
Compared to Claude Opus 4.6 ($15/$75), it is significantly cheaper, and close to GPT-5.4 Pro. With heavy cache usage, long-context task costs can be kept quite low.
It is currently in public preview, with a knowledge cutoff date of January 2025.
But It Is Not the Overall Leader
To be honest: on the overall benchmark ranking, GPT-5.4 Pro scores 92 (BenchLM.ai), Gemini 3.1 Pro scores 87, and Claude Opus 4.6 scores 85.
It won 12/18 individual tests, but still falls short in some key capability dimensions. In particular, for coding tasks, Claude Opus 4.6's SWE-bench performance remains stronger.
Google has delivered this report card on reasoning and multimodal processing, but overall leadership is not yet achieved.
Is It Worth Switching?
If your business primarily involves:
- Long document processing (contracts, reports, codebase reviews)
- Multimodal input (simultaneously processing images + text + audio)
- Limited budget but needing near-Opus-level reasoning
3.1 Pro is a serious option with competitive cost-effectiveness.
However, if the core need is writing code or building agents, Claude remains more mature. With this release, Google's 77.1% on ARC-AGI-2 shows that reasoning capability is genuinely improving, not just score-chasing. But to fully replace Claude and GPT-5.4, there is still some distance to cover.
Sources: Google's new Gemini Pro model has record benchmark scores — again (TechCrunch); CocoLoop, Gemini 3.1 Pro: A smarter model for your most complex tasks (Google DeepMind Blog); Gemini 3.1 Pro Preview Model Specs, Costs & Benchmarks (Galaxy.ai); Gemini 3.1 Pro | Google Cloud Vertex AI Documentation