Gemini 3.1 Ultra Hits 94.3% on GPQA Diamond, Setting New Reasoning Ceiling
Google DeepMind has completed the global rollout of Gemini 3.1, and the headline number is 94.3% on GPQA Diamond — a benchmark consisting of PhD-level science questions so difficult that non-experts typically score below 35% even with internet access. The previous record sat with GPT-5 internal builds. Gemini 3.1 Ultra is now the first publicly available model to exceed 94% on the eval.
What Changed Architecturally
The improvement is not a scaling story. Gemini 3.1 introduces a System 2 reasoning engine that runs a Chain-of-Verification (CoVe) process during inference. When the model encounters a high-complexity prompt, it generates internal sub-hypotheses, tests them against its own knowledge base and live search retrievals, and produces a final answer only after verification passes. Google says this reduces hallucination rates in technical and scientific outputs by over 60% compared to Gemini 2.0.
The Search-as-Logic integration is the key structural shift: Gemini 3.1 Ultra treats Google Search results as primitive inputs for a symbolic reasoning layer, not as retrieval-augmented context appended to a prompt. The distinction matters. Most RAG implementations inject retrieved text and let the model synthesize. Gemini 3.1 Ultra reasons over retrieved facts as typed logical primitives, which constrains the solution space before the answer token is generated.
Flash-Lite for Agentic Workloads
The Ultra gets the benchmark headlines, but the Flash-Lite variant is the production story. Google designed it specifically for agentic AI workflows where low latency and high throughput matter more than peak reasoning. Flash-Lite runs 2.5× faster than standard Flash at comparable output quality on most task categories.
For developers building agents that call tools, coordinate multi-step plans, or need sub-second response times at scale, Flash-Lite is the practical deployment target. The architecture supports 256K context, built-in thinking mode, vision understanding, and native function calling — the same capability surface as Ultra but tuned for speed.
Benchmark Context
GPQA Diamond has become the benchmark that separates genuine reasoning capability from pattern-matched plausibility. The trajectory of scores on it maps closely to real-world performance on hard scientific tasks:
| Model | GPQA Diamond |
|---|---|
| Gemini 3.1 Ultra | 94.3% |
| GPT-5.4 | ~91% (est.) |
| Claude Opus 4.6 | ~89% (est.) |
| Gemini 2.5 Pro | 78.4% |
| GPT-4o | 53.6% |
The gap from GPT-4o to Gemini 3.1 Ultra — 53.6% to 94.3% — covers roughly two years of frontier development. The gap from Gemini 2.5 Pro to Gemini 3.1 Ultra — 78.4% to 94.3% — represents a single model generation inside Google’s stack.
What It Means for the Leaderboard
Gemini 3.1 Pro Preview (the accessible API tier below Ultra) is already matching GPT-5.4 on Artificial Analysis’s Intelligence Index v4.0 at 57/126, while pricing 20% lower. If Ultra’s CoVe architecture propagates downward in future updates — as Gemini 3.0’s reasoning improvements did to Flash and Nano — the gap between frontier reasoning and production pricing compresses further.
The more significant implication is architectural: if verification-before-generation becomes the standard pattern at inference time, the hallucination rate gap between closed and open-weight models widens, because open-weight deployments cannot easily integrate live search as a logic primitive.