Google updates Gemini 2.5 Pro preview with improved coding performance

Google released an updated preview of Gemini 2.5 Pro, internally versioned 06-05, improving on the model’s performance in coding, mathematics, science and general reasoning. On LMArena’s crowdsourced leaderboard, the update gained 24 Elo points to reach a score of 1470, extending its lead at the top of the ranking; on WebDevArena, a leaderboard specifically for web-development tasks, it gained 35 points to reach 1443, also topping that board. The update led coding benchmarks including Aider Polyglot and performed well on GPQA and Humanity’s Last Exam, evaluations aimed at graduate-level science reasoning and broad expert knowledge respectively.

Google framed the release around demand rather than a single capability claim, describing Gemini 2.5 Pro’s usage growth as the steepest the company had seen for any of its models, and highlighting coding and agentic tasks as the areas where the update’s gains were most apparent to users. The update remained a preview release rather than a full production version, continuing the iterative-preview pattern Google had used throughout the Gemini 2.5 line since its initial release earlier in 2025.

The gains kept Gemini 2.5 Pro at the top of both LMArena and WebDevArena against competition from OpenAI’s o-series and Anthropic’s Claude 4 models, in a period when leading labs were updating flagship models every few weeks rather than on the multi-month cycles typical of 2023 and 2024 — a cadence that made any single leaderboard position transient.