On Thursday, Google unveiled an enhanced version of its Gemini 2.5 Pro artificial intelligence model. Titled “Gemini 2.5 Pro preview 06-05 Thinking,” the new iteration promises advanced coding capabilities compared to its predecessor, which debuted just before the Google I/O 2025 event. The company, based in Mountain View, noted that this update also refines the style and structure of its generated responses. This marks the second enhancement made to the model since its initial release.
Gemini 2.5 Pro Receives Further Enhancement
In a blog post, Google revealed plans for the new version to become widely available within weeks. Gemini 2.5 Pro was initially launched in March, receiving its first upgrade in May that focused on improving coding performance.
The latest version has shown significant progress on LMArena, a platform that ranks AI models based on user feedback. Google reported that the latest Gemini 2.5 Pro achieved a 24-point increase in its Elo score, now totaling 1,470. Additionally, it recorded a 35-point increase on the WebDevArena, reaching an Elo score of 1,443.
The company emphasized the model’s coding capabilities, noting its position as a leader on the Aider Polyglot benchmark, which is renowned for its challenging coding-related evaluations. Furthermore, Google stated that the Gemini 2.5 Pro maintains comparable scores on the GPQA and Humanity’s Last Exam (HLE) benchmarks.
In addition to the enhancements in performance, Google has also made strides in improving the style and organization of the model’s output. However, independent assessments to evaluate these improvements remain absent.
While the full, stable version is expected to be available in two weeks, the upgraded preview can currently be accessed through the Gemini app. It is also available via the Gemini application programming interface (API) on Google AI Studio and Vertex AI. Google has incorporated thinking budgets to allow developers enhanced control over cost and latency.
