Google just shipped Gemini 3.1 Pro, and it is not a minor version bump. The model arrives with a 2-million-token context window — the largest publicly available from any frontier lab — alongside native multimodal processing, a sandboxed code execution environment, and benchmark scores that put it at or near the top of every major leaderboard. If you have been waiting for the moment when a single prompt can swallow an entire codebase, a full-length film, or a 1,500-page regulatory filing, that moment is now.
What 2 Million Tokens Actually Buys You
Two million tokens translates to roughly 1,500 pages of dense text, several hours of video, or an entire medium-sized codebase loaded in one shot. Previous models forced developers to chunk documents, maintain retrieval pipelines, and stitch context together. Gemini 3.1 Pro eliminates that plumbing for a wide range of real-world tasks.
The context window is not a gimmick bolted onto an older architecture. Google DeepMind built the model around Ring Attention, a technique that divides ultra-long sequences into chunks, distributes them across multiple accelerators for parallel processing, and exchanges key information through an efficient communication ring. The result is near-linear scaling in computational complexity — the model does not choke when you push past the million-token mark.
Practical test: A 90-minute meeting transcript was summarized with structured action items in roughly 15 seconds, with accuracy close to 90%.
Benchmarks That Matter
Gemini 3.1 Pro's benchmark performance is hard to ignore:
| Benchmark | Gemini 3.1 Pro | Claude Opus 4.6 | GPT-5.4 |
|---|---|---|---|
| GPQA Diamond | 94.3% | ~88% | 92.8% |
| SWE-Bench Verified | 80.6% | 80.8% | — |
| ARC-AGI-2 | 77.1% | — | — |
| SWE-Bench Pro | 54.2% | — | — |
The GPQA Diamond score of 94.3% is the highest recorded on this graduate-level science reasoning benchmark. On SWE-Bench Verified, the standard coding benchmark, it hits 80.6% — neck and neck with Claude Opus 4.6 at 80.8%. And on ARC-AGI-2, the abstraction and reasoning test designed to resist memorization, it clears 77%.
These are not cherry-picked numbers. Google published results across 16 benchmarks, claiming wins on 13 of them. Independent researchers have noted that the remaining three — primarily creative writing and nuanced instruction-following — still favor Claude Opus 4.6.
Native Multimodal, Not Bolted On
Previous Gemini releases processed images and video by routing them through separate pipelines. Gemini 3.1 Pro handles text, image, audio, and video natively within the same architecture. You can upload a video and get summaries, timestamps, and action items without the model first transcribing the audio into text.
This matters for workflows like:
- Code review from screenshots: paste a UI screenshot alongside the component code, ask the model to find mismatches
- Video analysis: upload a product demo and get a structured breakdown of features demonstrated, with timestamps
- Audio + document cross-referencing: feed a recorded interview alongside a written transcript to flag discrepancies
Sandboxed Code Execution
Gemini 3.1 Pro ships with a Python sandbox built into the API. The model can write code, execute it, inspect the output, and iterate — up to five times per turn, with each execution capped at 30 seconds. This is not a novelty feature; it transforms the model from a code suggester into a code verifier.
Ask it to parse a CSV, and it will write the parsing script, run it, check the output, and fix edge cases — all within a single API call. For data science workflows, this eliminates the copy-paste loop between the model and a local environment.
Dynamic Thinking Levels
A new thinking_level parameter lets developers control how much chain-of-thought reasoning the model applies. Four settings — low, medium, high, and max — let you trade latency for accuracy depending on the task:
- Low: fast responses for simple lookups and formatting
- Medium: balanced for most business tasks
- High: deep reasoning for complex analysis
- Max: full chain-of-thought for research-grade problems
This is a meaningful UX improvement. Previous models either always thought deeply (slow and expensive) or never did (fast but shallow). Gemini 3.1 Pro lets you dial it per request.
Pricing and Access
Gemini 3.1 Pro is available through Google AI Studio, the Gemini API, and gemini.google.com for Advanced plan subscribers. API pricing sits at $2 per million input tokens and $12 per million output tokens. For prompts exceeding 200K tokens, the rate doubles to $4/$18 — still dramatically cheaper than competitors for long-context workloads.
A single 2-million-token input costs roughly $8 at the extended-context rate. Compare that to the per-token cost of feeding the same content through a competing model in multiple chunks, and the economics become compelling fast.
Where It Falls Short
No model dominates everything, and Gemini 3.1 Pro has clear gaps:
- Creative writing: Blind human evaluations in Q1 2026 preferred Claude Opus 4.6 output 47% of the time, versus 24% for Gemini 3.1 Pro
- Nuanced instruction-following: Tasks requiring careful adherence to complex, multi-step instructions still favor Claude
- Hallucination on niche topics: While grounding has improved, the model still occasionally fabricates details on obscure technical subjects
The 2-million-token context window is also not free from degradation. Performance on retrieval tasks drops measurably past the 1.5-million-token mark, though it remains usable for summarization and extraction.
The Bottom Line
Gemini 3.1 Pro is the most complete single-model offering available today. The 2-million-token context window is genuinely transformative for document-heavy workflows, the native multimodal processing eliminates pipeline complexity, and the sandboxed code execution turns it into a development partner rather than a suggestion engine. The aggressive pricing makes it viable for production workloads that were previously cost-prohibitive.
It is not the best model for every task — Claude still writes better prose, and the SWE-Bench coding crown belongs to Opus 4.6 by a hair. But for teams that need to process lots of stuff in one shot without breaking the budget, Gemini 3.1 Pro just set a new bar.
