Gemini 3.1 Pro: Google Takes the SWE-Bench Crown

Google released Gemini 3.1 Pro on February 19, two weeks after Anthropic’s Claude Opus 4.6. The headline benchmarks are strong: 80.6% on SWE-bench Verified, 94.3% on GPQA Diamond, and 77.1% on ARC-AGI-2. These numbers put Gemini 3.1 Pro at or near the top of every major benchmark category.

The SWE-Bench Number

SWE-bench Verified tests real software engineering ability. The model receives actual GitHub issues from popular repositories and has to produce code fixes that pass the associated test suites. Gemini 3.1 Pro’s 80.6% means it resolves more than four out of five of those issues.

To put this in context: Claude Opus 4.6, which released two weeks earlier, scored 80.8% on the same benchmark. Both numbers are the highest reported by any model at time of release, and they’re within a fraction of a percentage point of each other. The practical difference for most software engineering tasks is negligible. What matters is that both Anthropic and Google were now shipping models that resolve the large majority of benchmark issues without a human in the loop. That’s a real capability threshold, not a rounding error.

For developers building coding agents, these numbers matter. A model at 80% versus 50% on SWE-bench doesn’t just edge out the lower score. It means your agent handles a different class of problems on its own.

GPQA Diamond: 94.3%

GPQA Diamond is a graduate-level science reasoning benchmark. The questions require expert-level knowledge in biology, chemistry, and physics. These are the kind of problems doctoral students and subject-matter experts get right only about 65% of the time, and human expert performance overall sits around 69%.

Gemini 3.1 Pro at 94.3% is a striking number. It suggests the model has internalized a level of scientific reasoning that substantially exceeds typical expert human performance on the same questions. Previous leading scores on GPQA Diamond hovered in the high 70s to low 80s. A jump to 94.3% is significant.

This matters primarily for research, scientific analysis, and education use cases. For most developers, GPQA Diamond isn’t a day-to-day concern. But it signals something broader: the frontier models are moving past benchmark ranges that felt like hard ceilings six months ago.

ARC-AGI-2: 77.1%

ARC-AGI-2 is François Chollet’s benchmark for abstract visual reasoning: generalizing from a handful of examples to novel pattern-completion problems. It’s designed to measure a form of fluid intelligence rather than memorized knowledge. The original ARC-AGI benchmark was famously hard for early large language models; humans solve it at around 85%.

77.1% on ARC-AGI-2 is a strong result for a benchmark built to resist memorization. For reference, o3 with extended compute scored around 75% on ARC-AGI-1, which was considered a near-milestone performance. ARC-AGI-2 is designed to be harder, and Gemini 3.1 Pro clearing 77% on it is a data point worth tracking.

The ARC benchmark community is skeptical of results until replication is done by third parties. I’d expect those replications in the coming weeks.

Context and Multimodality

Gemini 3.1 Pro ships with a 1 million token context window (1,048,576 tokens) and a maximum output of 65,536 tokens. Claude Opus 4.6 also offered a 1 million token context window, so at launch the two frontier models were matched on context size rather than one holding a clear lead.

The model is natively multimodal, handling text, images, audio, and video in a single context. That helps for workflows that mix media: analyzing video alongside text descriptions, processing documents with embedded charts, or building applications that take in several input types at once.

Access and Pricing

Gemini 3.1 Pro is available through Google’s Gemini API and in AI Studio. Pricing follows a similar tier structure to earlier Gemini Pro models: $2 per million input tokens and $12 per million output tokens at standard rates, rising to $4 input and $18 output for requests over 200K tokens. At those rates it undercut Claude Opus 4.6 ($5 input / $25 output) at launch and stayed competitive with the OpenAI frontier models shipping at the time.

For organizations already in the Google Cloud ecosystem, Vertex AI integration means Gemini 3.1 Pro fits into existing infrastructure without additional vendor relationships.

The February 2026 Benchmark Moment

February 2026 is turning out to be a notable month for frontier model releases. Claude Opus 4.6 on February 5, then Gemini 3.1 Pro on February 19, with SWE-bench scores of 80.8% and 80.6% respectively. Both represent the highest-performing software engineering scores at time of release, and both clear 90%+ on the GPQA Diamond science reasoning benchmark.

The frontier is visibly bunching up now, with leading models from different labs landing at similar capability levels within weeks of each other. The era of one model staying clearly dominant for six months or more looks like it’s ending. For developers, that’s broadly good news: more competition, more choices, and a baseline capability floor that keeps rising.

That pace held up through the spring. By late May 2026, Google had made Gemini 3.5 Flash its new default model, claiming it surpasses Gemini 3.1 Pro on coding and agentic tasks, while Anthropic had moved on to Claude Opus 4.8. Gemini 3.1 Pro remains a capable Pro-tier option, but it is no longer the newest model in Google’s lineup.