Google is pushing its AI ambitions even further with Gemini 3, its most capable model to date, and one that aims to bring far more than smarter chat responses.
Announced alongside a sweeping rollout across Search, the Gemini app and its developer tools, the update marks Google’s biggest leap since the Gemini era began almost two years ago.
According to Google’s own research and internal testing, Gemini 3 is designed to understand deeper context, reason with more nuance and handle much more complex tasks with fewer prompts. CEO Sundar Pichai says the model brings Google “closer to AGI,” thanks to a combination of improved multimodal understanding, stronger agentic abilities and a major step up in raw reasoning power.
At the centre of the launch is Gemini 3 Pro, now in preview, which outperforms Gemini 2.5 Pro on every major benchmark. It currently tops the LM Arena leaderboard with an Elo score of 1501 and posts some striking results: 37.5% on Humanity’s Last Exam, Google’s internal tool-free reasoning test; 91.9% on GPQA Diamond; and a new state-of-the-art 23.4% on the MathArena Apex benchmark.
Google’s testing also shows big jumps in multimodal reasoning, with 81% on MMMU-Pro and 87.6% on Video-MMMU, plus improved factual accuracy at 72.1% on SimpleQA Verified.
Deep Think
A new Deep Think mode pushes these capabilities even further. In Google’s early evaluations, Deep Think delivers even higher scores, including 93.8% on GPQA Diamond and a frontier-level 45.1% on ARC-AGI-2 (with code execution). Google says it is withholding Deep Think from general release for additional safety review before making it available to Gemini Ultra subscribers.
Gemini 3 isn’t just aimed at chat-style interactions. It is built to help users learn, build and plan in more practical ways, using its larger 1-million-token context window and more advanced multimodal processing. That includes turning handwritten family recipes into a digital cookbook, analysing long academic videos and generating interactive study material, or even reviewing gameplay footage to produce training plans.
For developers, Gemini 3 arrives in Google AI Studio, Vertex AI, Gemini CLI and the new Google Antigravity platform – an agent-first development environment where Gemini can plan, execute and validate multi-step coding tasks. Google’s internal benchmarks show major jumps in agentic performance, including leading scores on WebDev Arena (1487 Elo), Terminal-Bench 2.0 (54.2%) and SWE-bench Verified (76.2%).
Coming to search
Gemini 3 is also beginning its rollout in Search via AI Mode, where Google is using the model to generate dynamic visual layouts, simulations and more contextual answers on the fly.
As always, Google emphasises that this is grounded in its own testing. The company says Gemini 3 is its most rigorously evaluated model yet, with reduced sycophancy, stronger defences against prompt injection and expanded external audits, including assessments from UK AISI, Apollo, Vaultis and Dreadnode.
Gemini 3 is available now in the Gemini app and across Google’s developer tools, with more models and features, including Deep Think, set to arrive in the coming weeks.
