Google’s Updated Gemini 3 Deep Think Outperforms GPT-5.2 and Claude Opus 4.6

Google, on Thursday, updates its Gemini 3 Deep Think artificial intelligence (AI) model. The frontier model was already the company's most intelligent model when it was launched in December 2025. Now, with this upgrade, Google says it can help scientists research challenging problems. The Mountain View-based tech giant highlighted that the update improves its performance across all major benchmarks, but most notably, the model sets new record on the ARC-AGI-2 and Humanity's Last Exam, outperforming both OpenAI's GPT-5.2 and Anthropic's Claude Opus 4.6.

Gemini 3 Deep Think Gets Upgraded

In a blog post, the tech giant said it is releasing a major upgrade to Gemini 3 Deep Think which will allow it to solve modern challenges across science, research, and engineering. The model continues to be available to the Google AI Ultra subscribers, but now, a select group of researchers and enterprises can also access it via the company's application programming interface (API).

Announcing the update, Google CEO Sundar Pichai said, “Gemini 3 Deep Think is getting a significant upgrade. We've refined Deep Think in close partnership with scientists and researchers to tackle tough, real-world challenges.” Elon Musk called the development “Impressive,” responding to the post.

With the improvement, the AI model is claimed to have scored 84.6 percent on the ARC-AGI-2 benchmark, which measures the reasoning capability of frontier models. Google claimed that the score was also verified by the ARC Prize Foundation. It also set a new record by scoring 48.4 percent (without tools) on Humanity's Last Exam, known for being the most difficult benchmark test in existence.

Additionally, the company claimed that Gemini 3 Deep Think also achieved Elo score of 3,455 on Codeforces. In each of these tests, the Google model is said to outperform frontier models from OpenAI and Anthropic.

Google also shared how some researchers are using the AI model in real-world scientific problems. It highlighted that Lisa Carbone, a mathematician at Rutgers University, used Gemini 3 Deep Think to review a highly technical mathematics paper. She observed that the model successfully identified a subtle logical flaw that had previously passed through human peer review unnoticed.