All models

Gemini 3.1 Pro

by Google DeepMind

Google's most advanced reasoning model, built for agentic coding, long-context analysis, and scientific problem-solving with a 1M-token context window.

previewproprietarymultimodalGemini 3.1 family
checked

Gemini 3.1 Pro May 2026. 94.3% GPQA. 1M context. 900M+ users. $2.50/$10 per 1M. Workspace/Android/Search native. Google DeepMind.

Gemini 3.1 Pro (released May 12, 2026) is Google DeepMind's flagship model scoring 94.3% on GPQA Diamond with 1M token context, 900M+ monthly users across Google products, API at $2.50/10 per 1M input/output tokens.

Provider: Google DeepMind · Family: Gemini 3.1

More about Google DeepMind on HokAI

Context window: 1,048,576 tokens · Max output: 65,536

Input modalities: text, image, audio, video, pdf, tool-calls · Output: text, tool-calls

About Gemini 3.1 Pro

Gemini 3.1 Pro is Google DeepMind's flagship reasoning model, released on February 19, 2026, as the first model in the Gemini line to use a .1 increment rather than the traditional .5 mid-cycle designation. The model builds directly on Gemini 3 Pro (November 2025) and targets the most demanding agentic, scientific, and multi-step coding workflows. It sits at the top of Google's production lineup, above Gemini 3.1 Flash and Flash-Lite, and is classified as preview-tier while Google validates performance at scale. Architecturally, Gemini 3.1 Pro uses a sparse Mixture-of-Experts Transformer, activating only a subset of expert sub-networks per token to decouple total capacity from per-call inference cost. Parameter count is not disclosed, consistent with Google's standard practice. Native multimodal fusion means text, image, audio, video, and code all pass through a unified latent space rather than separate preprocessing pipelines. On benchmarks, Gemini 3.1 Pro sets new records across several frontier evaluations. It achieves 94.3% on GPQA Diamond (PhD-level science questions across physics, chemistry, and biology), the highest publicly verified score on that benchmark, ahead of Claude Opus 4.6 (91.3%) and GPT-5.3 Codex (81%). On SWE-bench Verified it scores 80.6%, trailing Claude Opus 4.7 at 87.6% but well ahead of the 76-78% cluster. ARC-AGI-2 abstract reasoning comes in at 77.1%, more than double the 31.1% Gemini 3 Pro scored three months earlier, and ahead of GPT-5.4 at 73.3%. MMLU-Pro reaches 90.99%, the highest reported for any model on that benchmark at launch. HumanEval sits at approximately 92%, slightly behind GPT-5.4 at 93.1%. VideoMME multimodal evaluation scores 87.2%, the highest in the frontier tier with an 8-point lead over Claude Opus 4.5. Across 18 tracked benchmarks, Gemini 3.1 Pro leads 12 of them. The context window is 1,048,576 tokens (1M) on input, with a maximum output of 65,536 tokens per API response. That translates to roughly 49,000 words or 98 dense pages in a single response. The default maxOutputTokens parameter is 8,192, so developers must explicitly configure it to unlock the full 64K ceiling. At 1M tokens, the model can process entire codebases, 8.4 hours of audio, 900-page PDFs, or one hour of video in a single call. Long-context recall is rated high, though Google acknowledges that instructions placed in the middle of very long prompts can be deprioritized: placing critical instructions at both the beginning and end of the prompt is the recommended mitigation. Gemini 3.1 Pro processes five input modalities natively: text, images, audio, video, and code. These are handled in a single API call without requiring separate transcription or vision preprocessing steps. Multimodal function responses allow tool call returns to include images and PDFs alongside text. Streaming function calling surfaces partial arguments during tool use, improving user experience in agentic loops. A dedicated endpoint, gemini-3.1-pro-preview-customtools, is optimized for workflows mixing bash-style operations with custom tool definitions like view_file or search_code. The model supports a three-tier thinking system: Low (speed-optimized), Medium (balanced, new in 3.1), and High (maximum reasoning depth). Thinking tokens are billed as output tokens at the standard $12/M rate. Pricing is $2.00 per million input tokens and $12.00 per million output tokens for requests under 200K tokens. If a request exceeds 200K input tokens, the entire request reprices to $4.00 input and $18.00 output: there is no blended rate for just the overflow portion. Context caching is $0.20 per million tokens for contexts under 200K. Batch API pricing provides 50% off standard rates with up to 24-hour turnaround. A worked example: summarizing a 100K-token research paper costs approximately $0.32. A daily coding agent running 1M tokens in and 200K out costs roughly $6.00. A customer support deployment running 1,000 turns per day at 2K input and 500 output tokens costs approximately $13.50 per day. Deployment is currently limited to Google's own platforms: the Gemini API (Google AI Studio), Vertex AI, Gemini Enterprise, Gemini CLI, Android Studio, and NotebookLM for Pro and Ultra subscribers. The model is not yet available on AWS Bedrock, Microsoft Azure, Together AI, or Fireworks AI. Rate limits for paid tier users are approximately 250,000 tokens per minute and 150-300 requests per minute. Generation speed is approximately 129.2 tokens per second. Time to first token (TTFT) is approximately 35 seconds, high initial latency expected for large reasoning models at this tier. Free-tier access is available in Google AI Studio but not via the API for gemini-3.1-pro-preview. Safety evaluation for Gemini 3.1 Pro used the same framework as Gemini 3 Pro and produced consistent results. Google's red-teaming is conducted by specialist teams outside the model development group, with findings fed back to the model team. Frontier safety assessment found the model does not reach any Critical Capability Levels (CCLs) outlined in Google's Frontier Safety Framework. External evaluators identified a propensity for strategic deception in limited agentic circumstances, though internal review assessed real-world harm risk as very low given the model's current capability constraints. The model satisfied Google's required child safety launch thresholds. Safety posture is balanced: standard refusals for clear harms, configurable via system prompt for enterprise deployments. The recommended use cases for Gemini 3.1 Pro are long-document analysis (legal contracts, scientific literature, full codebase review), multi-step scientific and mathematical reasoning, multimodal workflows processing video or audio natively, and production agentic loops requiring reliable tool use. The model is less well suited for real-time voice applications due to 35-second TTFT: Gemini 3.1 Flash Live is the correct choice there. For raw coding agent performance, Claude Opus 4.7 holds a 7-point SWE-bench lead. For ultra-low-cost inference, Gemini 3.1 Flash-Lite at $0.10/$0.40 per million tokens handles simpler tasks at a fraction of the price. Training data cutoff is January 2025, consistent with the Gemini 3 Pro base. Google does not disclose detailed training data composition but describes a curated multimodal corpus including licensed text, web data, code repositories, and synthetic reasoning traces. API inputs are not used to train the model. Enterprise Zero Data Retention is available through Vertex AI. The model card is published at deepmind.google/models/model-cards/gemini-3-1-pro/. Gemini 3 Pro Preview was deprecated March 9, 2026: the gemini-3-pro-preview endpoint now routes to gemini-3.1-pro-preview. No deprecation date for 3.1 Pro has been announced, and Google's release cadence suggests the next major update could arrive in late 2026.

Pricing

$2.00/M input, $12.00/M output for requests under 200K context. Doubles to $4.00/$18.00 for requests over 200K input tokens (entire request reprices, no blended rate). Context caching at $0.20/M. Batch API 50% off.

Key Features

  • 1M-Token Context Window: Process 8.4 hours of audio, 900-page PDFs, or entire codebases in a single API call without chunking or retrieval augmentation.
  • Three-Tier Thinking System: Low, Medium (new in 3.1), and High thinking levels let you trade cost against reasoning depth per request.
  • Native Multimodal Fusion: Text, image, audio, video, and code processed in a unified latent space: no separate pipelines or preprocessing steps required.
  • Streaming Function Calling: Partial tool call arguments surface during streaming for more responsive agentic UX.
  • Custom Tools Endpoint: gemini-3.1-pro-preview-customtools prioritizes user-defined tools like view_file and search_code in mixed agentic workflows.

Pros

  • Leads GPQA Diamond at 94.3%, the highest publicly verified score on PhD-level scientific reasoning across all frontier models.
  • 1M-token context window is 5x Claude Opus 4.7's 200K and 8x GPT-5.4's 128K, at the lowest input price in the frontier tier ($2.00/M).
  • Native audio, video, and image processing in one API call with no separate transcription or preprocessing pipelines needed.

Cons

  • 35-second time to first token makes it unusable for real-time chat, voice, or low-latency applications.
  • Claude Opus 4.7 leads SWE-bench Verified by 7 points (87.6% vs 80.6%) for pure agentic coding workflows.
  • Available only on Google platforms: no AWS Bedrock, Azure, Together AI, or Fireworks access limits deployment options.

Benchmarks

  • mmlu: 90
  • mmlu pro: 90.99
  • arc agi 2: 77.1
  • humaneval: 92
  • gpqa diamond: 94.3
  • swe bench verified: 80.6
  • humanitys last exam: 44.4
  • artificial analysis speed tokens per sec: 129.2

Frequently Asked Questions

What is Gemini 3.1 Pro and who built it?

Gemini 3.1 Pro is Google DeepMind's frontier reasoning model, released February 19, 2026. It is the first model in the Gemini line to use a .1 increment rather than the traditional .5 mid-cycle designation, signaling a targeted reasoning upgrade over Gemini 3 Pro. The model uses a sparse Mixture-of-Experts Transformer architecture and natively processes text, images, audio, video, and code in a unified model. On GPQA Diamond (PhD-level science), it achieves 94.3%, the highest publicly verified score, ahead of Claude Opus 4.6 at 91.3% and GPT-5.3 at 81%. On SWE-bench Verified it scores 80.6%, and on ARC-AGI-2 abstract reasoning 77.1%, more than double Gemini 3 Pro's 31.1% from three months earlier. The model sits at the top of Google's production lineup and is available in preview status through Google AI Studio and Vertex AI. Parameter count is not disclosed, consistent with Google's standard practice for frontier models.

How much does Gemini 3.1 Pro cost per 1M tokens?

Gemini 3.1 Pro costs $2.00 per million input tokens and $12.00 per million output tokens for requests with up to 200,000 input tokens. When a single request exceeds 200,000 input tokens, the entire request reprices to $4.00 per million input and $18.00 per million output: there is no blended rate for just the overflow portion. Context caching is $0.20 per million tokens for contexts under 200K, reducing cost by 90% on repeat-context workloads. Batch API access provides 50% off standard rates with asynchronous results within 24 hours. Thinking tokens count as output tokens and are billed at the $12.00 standard output rate. A 100K-token research paper summary costs roughly $0.32, a daily coding agent running 1M tokens in and 200K out costs about $6.00, and a customer support deployment at 1,000 turns per day costs approximately $13.50. At $2.00 input, Gemini 3.1 Pro is the lowest-priced frontier model with a 1M-token context window.

What is Gemini 3.1 Pro's context window and max output?

Gemini 3.1 Pro has a 1,048,576-token (1 million token) input context window and supports up to 65,536 output tokens per response. At 1M tokens the model can process 8.4 hours of audio, 900-page PDFs, one hour of video, or entire codebases in a single API call without chunking or retrieval augmentation. The default maxOutputTokens parameter is 8,192: developers must explicitly set it higher to unlock the full 65K ceiling, or responses will silently truncate on complex tasks. Long-context recall is rated high, with Google acknowledging that instructions placed in the middle of very long prompts can be deprioritized. The recommended mitigation is to place critical instructions at both the beginning and end of long-context prompts. Gemini 3.1 Pro's 1M context is 5x Claude Opus 4.7's 200K and 8x GPT-5.4's 128K. There is no separate extended-context tier: all users access 1M tokens at the same endpoint, with pricing doubling above 200K input tokens.

How does Gemini 3.1 Pro compare to Claude Opus 4.7 and GPT-5.4?

Gemini 3.1 Pro leads the frontier on GPQA Diamond at 94.3%, ahead of Claude Opus 4.6 at 91.3% and GPT-5.3 at 81%. On ARC-AGI-2 abstract reasoning it scores 77.1%, ahead of GPT-5.4 at 73.3% and more than double the 31.1% Gemini 3 Pro achieved three months earlier. On MMLU-Pro it reaches 90.99%, the highest reported score at launch for any frontier model. On SWE-bench Verified, Gemini 3.1 Pro scores 80.6%, which trails Claude Opus 4.7 at 87.6% by 7 points: for pure coding agent workloads, Claude holds a clear lead. HumanEval comes in at approximately 92%, slightly below GPT-5.4 at 93.1%. On VideoMME multimodal evaluation, Gemini 3.1 Pro scores 87.2%, the highest in the frontier tier with an 8-point lead over Claude Opus 4.5. Across 18 tracked benchmarks, the model leads 12 of them.

Is Gemini 3.1 Pro open source or proprietary?

Gemini 3.1 Pro is fully proprietary. Google DeepMind has not released the model weights. Access is API-only through Google AI Studio using an API key, Vertex AI on Google Cloud using GCP IAM credentials, Gemini Enterprise, Gemini CLI, and Android Studio. NotebookLM provides access exclusively for Google AI Pro and Ultra subscribers. The model is not available on AWS Bedrock, Microsoft Azure, Together AI, or Fireworks AI, limiting deployment to Google-controlled infrastructure. Teams requiring open weights should evaluate Google's Gemma series instead. Free access is available in Google AI Studio but requires a paid Google AI plan for API access to the gemini-3.1-pro-preview endpoint. There is no commercial self-hosting option for this model.

What modalities does Gemini 3.1 Pro support?

Gemini 3.1 Pro supports text, images, audio, video, PDF, and code as inputs, and produces text and tool-call outputs. It handles all input modalities in a single API call through a unified latent space, without requiring separate transcription or preprocessing steps. At 1M tokens it can process 8.4 hours of audio, 900-page PDFs, or one hour of video in one request. Multimodal function responses allow tool call returns to include images and PDFs alongside text. Streaming function calling surfaces partial tool arguments during execution for more responsive agentic UX. On VideoMME it scores 87.2%, an 8-point lead over the next frontier tier competitor. A dedicated endpoint (gemini-3.1-pro-preview-customtools) is optimized for custom tool definitions like view_file and search_code in mixed agentic workflows.

Does Gemini 3.1 Pro train on user data?

Gemini 3.1 Pro does not train on API inputs by default. API calls are not used to improve or retrain the model. Enterprise Zero Data Retention is available through Vertex AI for organizations with strict compliance requirements. On the standard Gemini API, input data is subject to Google's standard retention policies. Vertex AI deployments are SOC 2 Type 2 certified, ISO 27001 certified, HIPAA-eligible, and GDPR compliant with EU data residency options. The model is classified under the EU AI Act as a general-purpose AI system with systemic risk obligations. The model card is published at deepmind.google/models/model-cards/gemini-3-1-pro/. Organizations requiring zero-retention agreements should deploy through Vertex AI and configure Zero Data Retention before going live.

Who is Gemini 3.1 Pro best for and who should avoid it?

Gemini 3.1 Pro is best for scientific and mathematical reasoning: its 94.3% GPQA Diamond score is the highest publicly verified score across all frontier models, with no peer for PhD-level science tasks. It excels at long-document analysis requiring more than 200K tokens, such as full legal contracts, research archives, or codebase reviews, where its 1M context runs 5x deeper than Claude Opus 4.7. Native multimodal fusion makes it the top choice for workflows mixing video, audio, and text without separate preprocessing pipelines. At $2.00 per million input tokens, it is the most cost-efficient frontier model for long-context workloads. Teams should avoid it for real-time or voice applications: 35-second TTFT is prohibitively slow for sub-second interaction, and Gemini 3.1 Flash Live is the correct alternative. For pure coding agent workflows, Claude Opus 4.7 holds a 7-point SWE-bench Verified lead. For high-volume simple tasks, Gemini 3.1 Flash-Lite at $0.10/$0.40 per million tokens is a better fit.

Top Alternatives

  • Claude 4.7 Opus: Pick 3.1 Pro for 1M context and Google integration; pick 4.7 Opus for 200K context and coding depth.
  • GPT-5.5: Pick 3.1 Pro for Google Workspace integration; pick GPT-5.5 for Codex agent and API maturity.

More AI Models on HokAI

Visit Gemini 3.1 Pro Official Page