[ad_1]
Google has kicked its Gemini rollout into high gear over the past year, releasing the much-improved Gemini 2.5 family and cramming various flavors of the model into Search, Gmail, and just about everything else the company makes.
Now, Google’s increasingly unavoidable AI is getting an upgrade. Gemini 3 Pro is available in a limited form today, featuring more immersive, visual outputs and fewer lies, Google says. The company also says Gemini 3 sets a new high-water mark for vibe coding, and Google is announcing a new AI-first integrated development environment (IDE) called Antigravity, which is also available today.
The first member of the Gemini 3 family
Google says the release of Gemini 3 is yet another step toward artificial general intelligence (AGI). The new version of Google’s flagship AI model has expanded simulated reasoning abilities and shows improved understanding of text, images, and video. So far, testers like it—Google’s latest LLM is once again atop the LMArena leaderboard with an ELO score of 1,501, besting Gemini 2.5 Pro by 50 points.
Factuality has been a problem for all gen AI models, but Google says Gemini 3 is a big step in the right direction, and there are myriad benchmarks to tell the story. In the 1,000-question SimpleQA Verified test, Gemini 3 scored a record 72.1 percent. Yes, that means the state-of-the-art LLM still screws up almost 30 percent of general knowledge questions, but Google says this still shows substantial progress. On the much more difficult Humanity’s Last Exam, which tests PhD-level knowledge and reasoning, Gemini set another record, scoring 37.5 percent without tool use.
Math and coding are also a focus of Gemini 3. The model set new records in MathArena Apex (23.4 percent) and WebDev Arena (1487 ELO). In the SWE-bench Verified, which tests a model’s ability to generate code, Gemini 3 hit an impressive 76.2 percent.
[ad_2]
Source link