Los Angeles: Google has introduced Gemini 3, its newest large language model and the company’s most capable AI system to date. The launch arrives only seven months after Gemini 2.5 and positions Google firmly in the race for cutting-edge AI development.

The release also comes less than a week after OpenAI released GPT 5.1, and a mere two months after Anthropic released Sonnet 4.5—a reminder of the blistering pace of frontier model development.

"With Gemini 3, we're seeing this massive jump in reasoning," said Tulsee Doshi, Google's head of product for the Gemini model. "It's responding with a level of depth and nuance that we haven't seen before."

Gemini 3 Deepthink pushes benchmark performance to new heights

Google plans to release Gemini 3 Deepthink, a more research-intensive version, to AI Ultra subscribers in the coming weeks, after it successfully completes further safety testing, TechCrunch reported.

Early evaluations show Gemini 3 Deepthink advancing well beyond Gemini 3 Pro’s strong performance in key benchmarks. It surpasses previous scores on Humanity’s Last Exam with 41.0% accuracy (without tool use) and achieves 93.8% on GPQA Diamond. The model also records an unprecedented 45.1% on ARC-AGI-2 with code execution, verified by the ARC Prize, reflecting its ability to solve novel challenges.

placeholder
Google Gemini 3 scores | Photo: Google

Learning, Building and Planning with Gemini 3

Gemini 3 is designed to support a wide spectrum of tasks. For learning, it can interpret handwritten family recipes from different languages and convert them into a digital cookbook. It can also digest academic papers, lengthy tutorials or video lectures and turn them into interactive study aids such as flashcards or visual explanations. The model can even review footage from activities such as pickleball games, highlight areas for improvement and create personalised training plans.

For developers, Gemini 3 is Google’s strongest agentic and vibe coding model to date. It leads the WebDev Arena leaderboard with a score of 1487 Elo and achieves 54.2% on Terminal-Bench 2.0, which measures the ability to operate a computer via terminal. It also performs exceptionally on SWE-bench Verified, scoring 76.2% and outperforming Gemini 2.5 Pro.

In everyday tasks, Gemini 3 can streamline complex, multi-step workflows—from organising inboxes to booking local services—by combining robust reasoning with improved tool use, all under the user’s direction.

According to Google, the Gemini app currently has more than 650 million monthly active users, and 13 million software developers have used the model as part of their workflow.

Google’s Antigravity elevates AI from assistant to active coding partner

Alongside the base model, Google also released a Gemini-powered coding interface called Google Antigravity, allowing for multi-pane agentic coding similar to agentic IDEs like Warp or Cursor 2.0.

Powered by Gemini 3’s enhanced reasoning, tool-use skills and agentic coding, Google Antigravity shifts AI from being a simple assistant to acting as an active collaborator. Although the platform resembles a conventional AI-enabled IDE, its integrated agents now occupy a dedicated workspace with direct access to the editor, terminal and browser. This setup allows them to independently plan and carry out complex, end-to-end software tasks, validating their own code as they progress.

Beyond support for Gemini 3 Pro, Google Antigravity is deeply integrated with Google’s latest Gemini 2.5 Computer Use model for browser interactions, along with Nano Banana (Gemini 2.5 Image), the company’s highest-rated image editing model.

(with inputs from ANI)