On February 19, 2026, Google DeepMind released Gemini 3.1 Pro Preview.
An improved version of Gemini 3 Pro, this model brings significant upgrades in reasoning depth, agentic capability, and long-context processing. Here's everything known as of today.
What Changed — Key Features of Gemini 3.1 Pro
1. Adjustable Thinking Level
The biggest new feature. You can now choose from three levels of Thinking depth: low, medium, and high.
- Low: Speed-first, lower cost, for simple tasks
- Medium: Balanced
- High: Called "Deep Think Mini" — deep reasoning mode for complex math, science, and coding
Previously it was binary: think or don't think. Now you can calibrate cost and depth to match the task.
2. Specs
- Context window: 1M tokens
- Max output: 64K tokens
- Input: text, audio, images, video, code
- Knowledge cutoff: January 2025
3. Improved Agentic Capability
Particularly stronger for multi-step task execution, tool calls, and autonomous coding workflows — use cases where you're running it as an agent.
Benchmark Comparison
Gemini 3.1 Pro vs major competing models.
Gemini 3.1 Pro (Thinking High) key scores:
- ARC-AGI-2 (abstract reasoning): 77.1% (Gemini 3 Pro: 31.1%, Opus 4.6: 68.8%)
- Humanity's Last Exam (academic reasoning): 44.4% (Opus 4.6: 40.0%)
- GPQA Diamond (scientific knowledge): 94.3% (GPT-5.2: 92.4%)
- SWE-Bench Verified (coding): 80.6% (Opus 4.6: 80.8%)
- Terminal-Bench 2.0 (terminal operations): 68.5% (GPT-5.3-Codex: 64.7%)
ARC-AGI-2 stands out. More than 2.5x the score of Gemini 3 Pro — a significant gap from competitors. SWE-Bench is effectively tied.
Use Cases
Coding and Development
An SWE-Bench score of 80.6% puts it in the range of delegating real production code fixes and implementations. Thinking High mode is especially useful for complex architecture design and difficult bug analysis.
Long-Context Processing
A 1M token window is strong for use cases where you feed in long specs, large codebases, or full papers and then ask questions.
Multimodal Tasks
Can process combined video, image, and audio input. UI design feedback, video summarization, audio transcript analysis.
Agents and Automation
Available via Google AI Studio, Vertex AI, and Gemini CLI. Well-suited for building autonomous workflows combined with agent frameworks like OpenClaw.
Where to Use It
- Google AI Studio (free trial available)
- Gemini API / Vertex AI
- Gemini CLI / Android Studio
- Gemini app (Google AI Pro/Ultra plans)
- NotebookLM, Gemini Enterprise
Migration Timing from Gemini 2.5 Pro
The stable version of Gemini 2.5 Pro is scheduled for retirement on June 17, 2026. Google recommends migrating to the 3.x series.
That said, this is still a Preview, so for production use it's safer to wait for the stable release. For testing and evaluation, it's usable now.
My Take
The 77.1% on ARC-AGI-2 was striking. Abstract reasoning more than double Gemini 3 Pro suggests something at the architecture level, not just fine-tuning.
The direction of "adjustable Thinking" is right. It's natural that deeper thinking costs more — being able to control that by task is useful for engineers.
Which to use vs. Claude depends on the task. For long context, multimodal, and agents, 3.1 Pro has an edge. For coding alone, it's roughly tied right now.
Curious about what Gemini 3 Ultra brings.
🐾
