Gemini 3 Pro Grabs the LMArena Lead and a New Humanity's Last Exam High
Google's new flagship ships on day one across Search, the Gemini app and the API, alongside an agentic coding tool called Antigravity.
Key takeaways
- Gemini 3 Pro scored 37.5% on Humanity's Last Exam without tools, 12.1 points above Grok 4.
- It topped LMArena at 1,501 points, ahead of Grok 4.1 Thinking at 1,483.
- Google launched it simultaneously in Search AI Mode, the Gemini app, the API and a new agentic IDE, Antigravity.
On November 18, 2025, Google released Gemini 3 Pro, calling it its most intelligent model yet. The launch came six days after OpenAI's GPT-5.1 and put Google at the top of the public leaderboards, with a record on one of the hardest reasoning tests around.
The numbers Google led with
On Humanity's Last Exam, a set of expert-written questions designed to stump frontier models, Gemini 3 Pro reached 37.5% without any tools such as web search. Engadget put that 12.1 points above the previous top score from xAI's Grok 4, which implies roughly 25.4% for Grok. Google's enhanced Deep Think mode pushed the figure to 41%, though it went first to safety testers and later to AI Ultra subscribers rather than shipping on day one.
On LMArena, the crowd-voted head-to-head ranking, Gemini 3 Pro landed at 1,501 points, ahead of Grok 4.1 Thinking at 1,483. It also led the WebDev Arena board for coding. Reports on OpenAI's score for HLE varied between outlets (one put GPT-5.1 at 26.5%), so the chart below sticks to the figures all outlets agreed on and labels the rest.
One model, many surfaces
The distribution was as notable as the scores. Gemini 3 Pro arrived at once in Search's AI Mode for Pro and Ultra subscribers, in the Gemini app under a "Thinking" option, and in the Gemini API, AI Studio and Vertex AI. Google said it would route harder AI Overviews queries to the model in the following weeks. A new Gemini Agent can take multi-step actions such as managing an inbox, with access to a user's Google apps.
Why it matters
Antigravity is the quiet strategic piece. It is an agentic coding environment where the model can plan tasks, write code and report progress on its own, which puts Google directly against Cursor-style editors and Claude Code. Shipping a frontier model and a developer product on the same day signals that Google now treats coding agents as a primary battlefield, not an add-on.
The timing matters too. Within a week of GPT-5.1 and a week before Anthropic's next release, the top spot on public leaderboards changed hands, which shows how short a lead now lasts. Arena-style rankings reward polish and style as well as correctness, so the 1,501 figure is a signal rather than a verdict.
How to read the leaderboard
Humanity's Last Exam was built to outlast easier tests. Its questions come from subject experts across disciplines and are designed so that leading models fail most of them. A score of 37.5% without tools means Gemini 3 Pro answered a little over a third correctly from its own knowledge and reasoning, which is notable precisely because the test was meant to stay out of reach. Still, benchmark creators publish only a portion of the questions, and labs optimise against public sets, so the long-run value of a record like this depends on whether it survives contact with held-out evaluations.
LMArena works differently. Users compare two anonymous answers and vote, and the platform converts votes into Elo-style ratings. A rating of 1,501 versus 1,483 is a narrow margin; the confidence intervals on such leaderboards often overlap for models in the same tier. The ranking captures how pleasant and persuasive answers feel, including formatting and tone, as much as strict correctness. That is useful for consumer products, where preference is the point, and less useful for judging reliability on a legal brief or a financial model.
The agent features carry their own caveats. Google's Gemini Agent needs permission to read and act inside a user's email and other Google apps, which concentrates risk: a model that mistakes an instruction embedded in an email for a command from the user could do real damage. Google positioned the agent as a gated capability for subscribers, and it will need to show that safeguards against prompt injection hold under attack. Antigravity raises the same issue for developers, since an autonomous coding agent with terminal access can delete files as easily as write them.
What to watch
- Whether Deep Think reaches Ultra subscribers on schedule and holds its 41% under independent testing.
- Developer uptake of Antigravity versus established coding assistants.
- How quickly OpenAI and Anthropic respond; the next flagships were already weeks away.
