Gemini 3 Pro Public Preview: Results From Early Access Evaluations

Gemini 3 Pro is the latest model in the Gemini family and represents a meaningful step forward over prior snapshots. During early access evaluations, the Harvey team observed stronger reasoning, cleaner structure, and more consistent style than prior Gemini models across our BigLaw Bench benchmark, product-focused evaluations, and unstructured lawyer usage. Below, we share a look into these results.

Gemini 3 Pro scored 87.9% on BigLaw Bench, an improvement over Gemini 2.5 Pro (85%) and in line with other frontier models. It particularly excelled at transactional work and in the litigation drafting sub-tasks, with better tone control, stylistic polish, and creative enhancements to content when appropriate.

In product‑oriented tests, Gemini 3 Pro was structured, stylistically consistent, and well‑formatted. It follows instructions tightly — including fine‑grained style requirements — and produces outputs with clear headers and correctly rendered tables. The model also performs especially well on long, multi‑item tasks, demonstrating robust reasoning over extended context.

Unstructured usage by our Applied Legal Research team points to meaningful progress in legal reasoning, drafting quality, and formatting discipline. Outputs are easy to navigate, with clear sections and logical flow. The model shows high instruction‑following across tone, formatting, and style, and adapts well to different drafting contexts.

We are working in parallel with the teams at Google and DeepMind to bring the model towards General Availability, at which point the model will be made available within Harvey. Stay tuned for more details!