Gemini 3.1 Pro Preview: what the .1 actually means

What shipped on February 11

Gemini 3.1 Pro Preview is not a new architecture. Google has described it as an improved checkpoint of the Gemini 3 series, with the same 1M token context window, the same multimodal inputs (text, images, audio, video, files), and the same pricing structure at $1.25 per million input tokens and $10 per million output tokens.

The changes Google documented are specific: improved instruction following on multi-step agentic tasks, stronger front-end code generation, better structured output fidelity, and updated video understanding weights trained on a wider distribution of video content. Those claims map, with varying precision, onto what the benchmarks actually show.

The word "preview" matters. This is not the general availability release. Google has historically used preview periods of six to twelve weeks before locking a checkpoint for production SLA guarantees. Teams running production workloads should benchmark before replacing Gemini 3 Pro.

The increment: Gemini 3 Pro to 3.1 Pro Preview

The cleanest way to evaluate a minor version bump is the delta, not the absolute scores. Here is every benchmark we tracked across the two checkpoints.

The pattern is consistent: every benchmark moved up between 1.2 and 5.2 percentage points. No regressions appeared in our testing. That said, the gains are not uniform. Math and coding improved most sharply. MMLU, which was already strong, saw the smallest lift. VideoMME moved meaningfully despite being a harder benchmark to shift, which suggests the video weight update was substantive rather than cosmetic.

Reasoning and math: where it leads the frontier

AIME 2025 at 91.2% puts Gemini 3.1 Pro Preview ahead of every proprietary competitor in pure mathematical reasoning. Claude Opus 4.5 scores 84.5% on the same benchmark. GPT-5.1 sits at 82.1%. The gap over Claude is 6.7 percentage points, which is not noise.

The one model that still beats it on math is DeepSeek R2 at 93.8%. DeepSeek R2 is an open-weight reasoning specialist built specifically around chain-of-thought depth. It does not process images, audio, or video. Comparing them on AIME is like comparing a sprinter to a triathlete on the hundred-meter split: accurate, but incomplete.

Coding: real improvement, real gap remaining

SWE-Bench Verified is the coding benchmark that most closely approximates real software engineering work. It asks models to resolve GitHub issues from real open-source repositories, with external agents able to use tools. A higher score means more issues resolved autonomously.

Gemini 3.1 Pro Preview scores 71.4% on SWE-Bench Verified. That is a 4.2 percentage point improvement over Gemini 3 Pro. It is also 4.9 percentage points below Claude Opus 4.5, which scores 76.3%.

For coding-first teams, that gap matters. In a 100-issue backlog, Claude Opus 4.5 resolves roughly 5 more issues autonomously. Over months of agentic workflows, that compounds. The counterargument is cost: Claude Opus 4.5 bills at $15 per million input tokens and $75 per million output tokens. Gemini 3.1 Pro Preview bills at $1.25 input and $10 output. At 12x the price difference, you can run a lot of Gemini iterations for the same budget.

Where Gemini 3.1 Pro Preview holds a real edge in coding is front-end work. Its WebDev Arena rank remains #1 at an Elo of 1,443. For generating React components, writing CSS, creating full-page UI from descriptions, and transforming design mockups to code, it is the strongest option available. Claude Opus 4.5 is not far behind, but Gemini's multimodal training gives it an edge when the input includes screenshots, Figma exports, or visual references.

Video understanding: the clearest improvement

VideoMME is the benchmark where Gemini 3.1 Pro Preview makes its most defensible claim to leadership. At 87.2%, it outperforms every other model in this comparison by a meaningful margin. Claude Opus 4.5 scores 79.2%. GPT-5.1 scores 81.3%. DeepSeek R2 does not support video at all.

The VideoMME benchmark tests comprehension, temporal reasoning, and question answering across video clips of varying lengths. The 2.4 percentage point gain over Gemini 3 Pro is explained by an updated video encoder trained on longer-form content. Google has confirmed that the model now reliably handles clips up to three hours, whereas Gemini 3 Pro degraded noticeably past the 90-minute mark.

For product teams building on video, this matters practically. Podcast summarization, meeting transcription and analysis, video search, video-to-code workflows: all of these improve with a stronger VideoMME baseline. The gap over Claude Opus 4.5 on video is nearly 8 percentage points, which is the largest differential across any benchmark in this comparison.

The full field: all models, all benchmarks

The table reveals a clear pattern: no single model wins every category. DeepSeek R2 leads on AIME but is absent from multimodal benchmarks. Claude Opus 4.5 leads on SWE-Bench and MMLU but trails badly on video. Gemini 3.1 Pro Preview wins four of six columns while maintaining the top LM Arena position, which aggregates human preference across all task types.

Where it still loses

Honest benchmark coverage requires calling out the categories where Gemini 3.1 Pro Preview is not the answer.

The pricing argument

Pricing in this tier has not converged. The spread between models is still 27x from cheapest to most expensive for comparable capability levels. That spread matters for volume applications.

At volume, the Gemini 3.1 Pro Preview pricing is genuinely disruptive. Running a 10-million-token-per-day pipeline costs $56 per day on Gemini vs $450 per day on Claude Opus 4.5. That is $394 per day or roughly $143,000 per year saved per 10M token pipeline. For enterprises running multiple products, that margin difference changes build-vs-buy calculus on entire features.

The 1M token context window also makes the pricing case stronger for long-context applications. Claude Opus 4.5 caps at 200K tokens, meaning long-document processing requires chunking, retrieval engineering, or multiple calls. Gemini 3.1 Pro Preview handles up to 1M tokens natively, eliminating that architectural complexity for many use cases.

Context utilization: the 1M token question

A 1M token context window is only as useful as a model's ability to retrieve information from it accurately. Long context benchmarks measure not just whether a model accepts long inputs but whether it uses them correctly.

Gemini 3.1 Pro Preview scores 93.4% on RULER (the long-context retrieval benchmark), compared to 91.2% for Gemini 3 Pro. For reference, Claude Opus 4.5 scores 94.1% at its 200K limit. The practical implication is that Gemini 3.1 Pro Preview reliably retrieves information from long documents, but is not yet flawless. In our testing with 800K-token legal contract sets, the model missed one relevant clause in approximately every 15 queries. That is significantly better than most alternatives but worth testing on your specific document distribution before committing.

Who should actually switch

The vendor concentration question

One factor that benchmarks do not capture is the concentration risk of going deeper into any single provider. Teams already heavily dependent on Google Cloud services gain workflow integration benefits from Gemini. Teams on AWS or Azure have reasons to evaluate the GPT-5.1 path even at higher per-token cost, since native integrations and unified billing often offset API savings.

Google has also shown a willingness to adjust model pricing mid-cycle, as it did between the Gemini 2.5 and Gemini 3 generations. The current $1.25/$10 pricing may not hold for the GA release of Gemini 3.1 Pro. Budgeting based on preview pricing carries some risk.

The verdict

Benchmark data sourced from Google DeepMind technical report (February 2026), Artificial Analysis, LM Arena, and independent SWE-Bench evaluations. Pricing verified against Google AI Studio and Vertex AI as of February 20, 2026. See our interactive model comparison for the latest pricing and benchmark data across 100+ models.