Gemini 3.1 Pro Preview
Google's frontier reasoning model, leading on novel logic (ARC-AGI-2 77.1%), graduate-level science (GPQA Diamond 94.3%) and agentic reliability, with a 1M-token context and full multimodal input — built for complex software engineering, long-document analysis and autonomous agent workflows.
- Modalities
- Chat
- Starting price
- From $0.15 / call
- Context
- 1M context
README
Gemini 3.1 Pro Preview is a native multimodal reasoning model in Google DeepMind's Gemini 3 series, released on February 19, 2026, under the model code gemini-3.1-pro-preview. It is currently in Preview. Based on Gemini 3 Pro, it can jointly understand text, images, video, audio, PDFs, and code repositories within a 1,048,576-token context window and produce up to 65,536 text tokens. Google has not disclosed its parameter count.
Compared with Gemini 3 Pro, this release focuses on better thinking, token efficiency, factual consistency, software-engineering usability, and reliable multi-step agent execution. In Google's February 2026 evaluations, the high-thinking configuration scored 77.1% on ARC-AGI-2, 80.6% on SWE-Bench Verified with a single attempt, and 68.5% on Terminal-Bench 2.0, all substantially above Gemini 3 Pro. It also supports function calling, structured outputs, code execution, context caching, search grounding, Google Maps grounding, and URL context.
Key Capabilities
- Complex Reasoning: Uses configurable Thinking for abstract logic, scientific knowledge, and multi-stage problems, scoring 77.1% on ARC-AGI-2 and 94.3% on GPQA Diamond in Google's evaluations.
- Advanced Coding: Optimized for software engineering, algorithm development, and repository-scale work, with 80.6% on SWE-Bench Verified and a 2887 LiveCodeBench Pro Elo.
- Agents and Tool Use: Supports function calling, code execution, search, and structured outputs, scoring 68.5% on Terminal-Bench 2.0 and 69.2% on MCP Atlas.
- Native Multimodal Understanding: Jointly analyzes text, images, video, audio, and PDFs for complex documents, visual material, recordings, and cross-modal evidence.
- Million-Token Long Context: Handles large codebases, long document collections, and cross-file project material within a 1,048,576-token input window, with context caching support.
- Retrieval and Factual Grounding: Can use Google Search, Google Maps, and URL context to incorporate current external information and improve verifiability.
- Controlled Structured Output: Produces structured formats such as JSON and combines them with function calling for integration into business systems and automated workflows.
Technical Strengths
| Feature | Benefit |
|---|---|
| Native Multimodal Reasoning | Connects text, visual, audio, video, and document information in one context, reducing semantic loss from separate processing pipelines. |
| 1M-Token Context | Processes large repositories or multi-document corpora in one request, reducing manual chunking, summary chaining, and missed dependencies. |
| Enhanced Thinking | Allocates additional reasoning effort to difficult tasks, improving abstract reasoning, scientific Q&A, and multi-step problem solving. |
| Precise Tool Use | Function calling, code execution, search grounding, and structured outputs work together to make agents more reliable in real workflows. |
| Improved Token Efficiency | Uses tokens more effectively than Gemini 3 Pro on complex work, helping manage cost and context across long-running processes. |
| Dedicated Custom Tools Endpoint | gemini-3.1-pro-preview-customtools prioritizes bash and custom tools for code browsing and tool-intensive agent workflows. |
Pricing
Tiered by input prompt tokens (incl. cache): once over the threshold, the whole request is billed at the higher tier.
| Token Type | Short context ≤200K | Long context >200K |
|---|---|---|
| Input | $1.5$2-25% | $3$4-25% |
| Output | $9$12-25% | $13.5$18-25% |
| reasoning_tokens | $9$12-25% | $13.5$18-25% |
| Cached input | $0.15$0.2-25% | $0.3$0.4-25% |