Gemini 3.1 Pro
Google's deep-reasoning flagship — #1 of all models on GPQA Diamond (94.1%) and 77.1% on ARC-AGI-2, with full multimodal input across a 1M-token context. In public preview since February 19, 2026, from $2/$12 per million tokens.
Model Specs
- Released
- Feb 2026
- Context window
- 1.0M tokens
- Max output
- 66K tokens
- Capabilities
- reasoningmultimodalfunction-callinglong-context
- Modalities
- textvisionaudiovideo
About this model
Gemini 3.1 Pro is Google's deep-reasoning flagship, released to public preview on February 19, 2026 — where it debuted at #1 on the Artificial Analysis Intelligence Index before the spring wave of releases (Opus 4.8, GPT-5.5, Fable 5) reshuffled the leaderboard. It still holds the single most impressive specialist crown: **94.1% on GPQA Diamond, the highest of any model**, plus 77.1% on ARC-AGI-2 (ARC-verified) — the strongest published score on the fluid-intelligence benchmark — and 44.7% on Humanity's Last Exam.
The profile is scientific depth plus multimodality. SciCode 58.9% (#2 of all models), CritPt 18% (led at launch), MMMU-Pro 82% — and unlike most frontier rivals it accepts text, images, video, and audio natively in a 1,048,576-token context. On AA-Omniscience it cut its hallucination rate by 38 points versus Gemini 3 Pro, ranking #2 on the honesty index behind only Claude Fable 5. Output streams at ~110 tokens/sec with a 25-second first token — notably snappier than the 100s+ deliberation of GPT-5.5 or Fable 5 at similar reasoning depth.
Pricing is tiered by context: $2/M input and $12/M output up to 200K tokens, doubling to $4/$18 beyond, with $0.20/M context caching. One honest caveat: as of mid-June 2026 the model remains in **public preview** (gemini-3.1-pro-preview) — fully priced and broadly usable via the Gemini API, but without a stable GA identifier yet; Google's announced Gemini 3.5 Pro will eventually take this slot. Reach for Gemini 3.1 Pro when scientific reasoning, abstract problem-solving, or video/audio understanding matter most — and consider Gemini 3.5 Flash when speed and price beat depth.
Key Strengths
#1 on GPQA Diamond — of all models
94.1% on graduate-level science reasoning, the highest score of any model on the Artificial Analysis harness — ahead of Opus 4.8 (93.6%), GPT-5.5 (93.5%), and Fable 5 (92.6%).
Best published ARC-AGI-2 score
77.1% (ARC-verified) on the fluid-intelligence benchmark designed to resist memorization — well clear of Gemini 3.5 Flash's 72.1% and unpublished by most rivals.
Full multimodal input
Text, images, video, and audio natively in one request — analyze recorded meetings, lecture videos, or screenshots alongside code and documents. Most frontier rivals are text+image only.
Scientific reasoning depth
SciCode 58.9% (#2 overall) and CritPt 18% (led at launch) — research-grade performance on scientific computing and frontier physics reasoning.
Honesty gains
Hallucination rate down 38 points versus Gemini 3 Pro; #2 on the AA-Omniscience honesty index behind only Claude Fable 5.
Reasonable latency for its depth
~110 tokens/sec with a 25s first token — a fraction of the 100s+ wait of GPT-5.5 or Fable 5, making deep reasoning usable in semi-interactive flows.
How it compares
Gemini 3.1 Pro is the science-and-multimodality specialist of the frontier tier — compare on reasoning profile, modality needs, and price.
| vs. Model | Verdict | Outcome |
|---|---|---|
| Gemini 3.5 Flash | Different jobs: 3.5 Flash is newer, faster (152 vs 110 t/s, 18.7s vs 25s TTFT), cheaper ($1.50/$9 flat), and actually beats 3.1 Pro on coding/agentic benchmarks. 3.1 Pro wins on hard reasoning (HLE 44.7 vs 41.0, ARC-AGI-2 77.1 vs 72.1, GPQA #1). Flash for production agents; 3.1 Pro for science and depth. | Depends |
| GPT-5.5 | GPT-5.5 leads the composite Intelligence Index (60.2 vs 57) and real-work GDPval, but 3.1 Pro takes GPQA Diamond (94.1 vs 93.5), ARC-AGI-2, and adds video/audio input at a lower price ($2/$12 vs $5/$30 under 200K). Science and multimodality favor Gemini; agentic professional work favors GPT-5.5. | Depends |
| Claude Opus 4.8 | Opus 4.8 leads composite intelligence (61.4 vs 57), real-codebase engineering (SWE-Bench Pro 69.2% vs 54.2%), and offers 1M context with no pricing tier jump. 3.1 Pro counters with GPQA #1, ARC-AGI-2, video/audio input, and lower entry price. Code and trust → Opus; science and multimodal → Gemini. | Other wins |
Pros
- #1 of all models on GPQA Diamond (94.1%)
- Best published ARC-AGI-2 score (77.1%, ARC-verified)
- Native video + audio + image input in 1M context
- SciCode 58.9% (#2) — research-grade scientific computing
- Hallucination rate down 38 points vs Gemini 3 Pro
- Lower entry price than GPT-5.5/Opus 4.8 ($2/$12 under 200K)
- Faster to first token (~25s) than deep-reasoning rivals (~100s+)
Things to consider
- Still in public preview — no stable GA model id as of mid-June 2026
- Pricing doubles above 200K context ($4/$18)
- Trails Opus 4.8/GPT-5.5/Fable 5 on the composite Intelligence Index (57)
- Coding/agentic benchmarks now beaten by its own cheaper sibling (3.5 Flash)
- 64K output cap — half of Opus/GPT-5.5's 128K
- Gemini 3.5 Pro is announced and will eventually supersede it
Best use cases
Scientific research & analysis
The GPQA Diamond and SciCode leader: literature analysis, hypothesis exploration, research-grade scientific computing.
Video & audio understanding
Native video/audio input in 1M context — meeting analysis, lecture summarization, media QA that text-only frontier models simply can't do.
Novel problem solving
The ARC-AGI-2 record reflects genuine fluid intelligence — useful for problems with no template: novel algorithms, unusual constraints, greenfield design.
Long-document multimodal synthesis
Mix PDFs, slides, recordings, and notes in a single 1M-token request for cross-source synthesis.
Math & competition reasoning
Deep-reasoning profile suits olympiad-style math, formal logic, and proof sketching.
High-trust drafting
The 38-point hallucination reduction makes it a credible drafting partner for technical and factual content.
How to use it on Renas AI
- 1
Step 1
Open AI Chat on Renas
Navigate to AI Chat in the Renas dashboard and pick the most capable Gemini model available to your plan.
- 2
Step 2
Bring multimodal evidence
Attach the video, audio, or images directly instead of describing them — native multimodal input is the differentiator over text-only rivals.
- 3
Step 3
Ask the hard science questions
This is the GPQA leader: give it graduate-level problems, full derivations, and research questions rather than saving it for routine chat.
- 4
Step 4
Watch the 200K pricing tier
Input beyond 200K tokens doubles the rate ($2→$4 in, $12→$18 out). For huge contexts at flat pricing, Claude Opus 4.8 may be more economical.
Pricing
Pricing on Renas AI
Pay-as-you-go credits, no API keys, no rate limits.
~142,857 words on a 10,000-credit Spark plan
Frequently asked questions
Other Google models
Other text models on Renas AI
Google's deep-reasoning flagship on Renas
Access leading Gemini models with your Renas AI subscription credits — no API key, no setup, no per-seat fees.
Try Gemini models on Renas