Google Releases Upgraded Preview of Gemini 2.5 Pro, Better at Coding and Reasoning

Google has launched an upgraded preview of Gemini 2.5 Pro, which will become the stable, generally available version in the coming weeks. The update builds on the version introduced in May and presented at the Google I/O conference, and is intended for enterprise-scale use.

The new model maintains its lead on multiple benchmarks. It shows a 24-point Elo increase on LMArena, taking its score to 1,470, and a 35-point jump on WebDevArena, now standing at 1,443. 

“The latest 2.5 Pro reflects a 24-point Elo score jump on LMArena…and a 35-point Elo jump to lead on WebDevArena,” said Tulsee Doshi, senior director and product lead, on behalf of the Gemini team.

Gemini 2.5 Pro also leads in coding benchmarks, including Aider Polyglot. It ranks high on academic benchmarks such as GPQA and Humanity’s Last Exam, which test a model’s capabilities in mathematics, science, knowledge and reasoning.

The team addressed developer feedback on earlier iterations of 2.5 Pro, improving response formatting and output structure. “We also addressed feedback from our previous 2.5 Pro release, improving its style and structure — it can be more creative with better-formatted responses,” said Doshi.

Developers can access the upgraded preview through the Gemini API via Google AI Studio and Vertex AI. Google has also introduced thinking budgets in these tools to help developers manage cost and latency. The model is also being rolled out in the Gemini app starting June 5.

At Google I/O, the tech giant also introduced Deep Think for the Gemini 2.5 Pro models, further enhancing their reasoning capabilities. With Deep Think, the model scored 49.4% on the United States of America Mathematical Olympiad (USAMO), outperforming the Gemini 2.5 Pro without Deep Think (34.5%) and OpenAI’s o3 (21.7%) and o4-mini (19.2%) models.

In addition, Google unveiled Gemini Diffusion, a model that applies methods from image and video generation to produce outputs, with a reported processing speed of 1,479 tokens per second."