Gemini 3.1 Pro Model Card
by Google DeepMind
The primary source for what Gemini 3.1 Pro can actually do — and where it hit a safety alert threshold.
Overview
Published 19 February 2026, the Gemini 3.1 Pro model card is Google DeepMind's official specification sheet for the next iteration in the Gemini 3 series of natively multimodal reasoning models. It follows the standard model-card structure: model description and architectural lineage from Gemini 3 Pro, inputs and outputs, training data and process, evaluation results, safety evaluation, and intended usage and limitations. The hard numbers an engineer needs are all here — text, image, audio and video input, a context window of up to 1 million tokens, and a 64K-token output ceiling. The evaluation section reports results across reasoning, multimodal, agentic tool-use, multilingual and long-context suites, including Humanity's Last Exam at 51.4% with tools, ARC-AGI-2 at 77.1%, GPQA Diamond at 94.3%, SWE-Bench Verified at 80.6% and a LiveCodeBench Pro Elo of 2887. The safety section reports testing under Google DeepMind's Frontier Safety Framework across five critical-capability domains: the model remains below alert thresholds for CBRN, harmful manipulation, machine-learning R&D and misalignment, but reaches the alert threshold for cyber capabilities — the single most decision-relevant line for anyone deploying it in a security-sensitive context. The card also enumerates distribution surfaces: the Gemini app, Gemini API, Google AI Studio, Google Cloud and Vertex AI, Google Antigravity, Gemini Enterprise and NotebookLM.
At a Glance
- Topic
- Models
- Level
- Intermediate
- Format
- Documentation
- Cost
- Free
- Duration
- ~25 min read
- Provider
- Google DeepMind
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓Read a frontier model card's structure: capabilities, evaluations, safety and stated limitations
- ✓Cite Gemini 3.1 Pro's exact context window, output ceiling and supported input modalities
- ✓Compare reported scores across reasoning, coding, agentic tool-use and long-context benchmarks
- ✓Understand how Frontier Safety Framework critical capability levels and alert thresholds are reported
- ✓Identify which Google surfaces expose the model to enterprise and API workloads
- ✓See how first-party benchmark reporting diverges from independent third-party evaluations
Highlights
- •A primary source you can quote, rather than a benchmark roundup that has already drifted
- •Discloses that the cyber-capability alert threshold was reached — the kind of line vendor marketing pages omit
- •Gives one dated, self-contained snapshot suitable for a procurement or AI risk review
- •Third-party write-ups already contradict it on context window and ARC-AGI-2 figures, which is exactly why you read the card
- •Covers consumer, API and Vertex AI / Gemini Enterprise deployment surfaces in one place
Who It’s For
Best For
- ✓Engineers choosing between frontier models on measured capability rather than marketing
- ✓AI risk, security and governance reviewers who need a citable primary source
- ✓Analysts and technical writers tracking release-to-release model capability deltas
Prerequisites
- •Familiarity with common LLM benchmarks such as GPQA Diamond, SWE-Bench and ARC-AGI
- •Basic understanding of tokens, context windows and multimodal model inputs
FAQ
What is Gemini 3.1 Pro Model Card?
Google DeepMind's official model card for Gemini 3.1 Pro, published February 2026. It is the citable primary source for the model's modalities, context window, benchmark scores and Frontier Safety Framework results, and it is what you should quote in an architecture review or vendor evaluation instead of a secondhand benchmark roundup that has already drifted from it.
Is Gemini 3.1 Pro Model Card free?
Gemini 3.1 Pro Model Card is free to access.
What level is Gemini 3.1 Pro Model Card for?
Gemini 3.1 Pro Model Card is aimed at a intermediate audience. Recommended background: Familiarity with common LLM benchmarks such as GPQA Diamond, SWE-Bench and ARC-AGI, Basic understanding of tokens, context windows and multimodal model inputs.
How long does Gemini 3.1 Pro Model Card take?
Expect roughly ~25 min read. Most learners work through it at their own pace.
What will I learn from Gemini 3.1 Pro Model Card?
You'll learn: Read a frontier model card's structure: capabilities, evaluations, safety and stated limitations; Cite Gemini 3.1 Pro's exact context window, output ceiling and supported input modalities; Compare reported scores across reasoning, coding, agentic tool-use and long-context benchmarks; Understand how Frontier Safety Framework critical capability levels and alert thresholds are reported; Identify which Google surfaces expose the model to enterprise and API workloads; See how first-party benchmark reporting diverges from independent third-party evaluations.
Topics
Sources
This page was written from 2 sources, 1 on domains other than deepmind.google.