Gemini 3.1 Pro Review: The Reasoning Leader You Haven't Tested

Gemini 3.1 Pro just scored 77.1% on ARC-AGI-2 — more than double its predecessor, and 24 percentage points above GPT-5.5's 52.9% on the same benchmark. ARC Prize confirmed the result in February 2026, and on the benchmark designed specifically to resist memorization and reward fluid reasoning, no other frontier model is close. If you have been picking models off the LMArena board, you have been making this decision without that number in front of you.

The Elo is the reason. Gemini 3.1 Pro sits at 1493 on Arena, sandwiched between Claude Opus 4.7 at 1503 and GPT-5.5 at 1484. That ten-point spread looks like a near-tie, and on the workloads Arena actually measures — open-ended chat, coding turns, single-shot tasks judged by humans — it more or less is. The problem is that Arena aggregates preference votes across millions of casual prompts, which compresses the signal on the long-tail reasoning problems where a model either gets it right or doesn't. A 24-point ARC-AGI-2 gap and a 9-point Elo gap can coexist because they are measuring different things, and the leaderboard most readers anchor on hides the larger of the two.

GPQA Diamond tells the same story from a different angle. Gemini 3.1 Pro leads at 94.3% — the highest score recorded on the graduate-level science benchmark — edging out Claude Mythos Preview's 93.9% on a different eval set. Together with ARC-AGI-2, that is two independent reasoning evals where Gemini is the unambiguous leader, not the third-place model the Arena ranking implies. For research workloads, multi-step scientific question answering, and any task where the bottleneck is deduction rather than fluency, the case for Gemini is no longer a tie-breaker. It is the default.

The honest counterweight is that reasoning benchmarks do not automatically translate into production coding wins. Claude Opus 4.7's Arena lead is real for the workloads it reflects — agentic coding loops, repo-scale refactors, long-running tool use — and developers running those flows have been consistent about why they pay the premium. Gemini at $2 input and $12 output per million tokens is the cheapest of the four tracked frontier models on input and 58% cheaper on output than Opus, but price comparisons stop mattering when a model abandons branches or edits the wrong file. Until the SWE-Bench and agentic coding numbers move, Opus stays the right call for those workloads.

What the ARC-AGI-2 result actually signals is the direction of travel. Released in preview on February 19 with a 1M context window matching the field, Gemini 3.1 Pro has reasoning headroom that the other frontier models do not. If Google closes even half the agentic-coding gap in the next release — and a reasoning leader usually does — the price advantage and the benchmark spread compound into the obvious default for mixed workloads. The model worth testing this quarter is the one most teams have not gotten around to.