Google has announced a significant update for its advanced AI model, Gemini 2.5 Pro. The latest version, named "06-05," aims to address issues identified in a previous update. This release is being promoted as a stable version that can be accessed through the Gemini app.
Performance Boost
The previous iteration of the Gemini 2.5 Pro model, also referred to as the I/O Edition or 05-06, primarily focused on enhancing coding capabilities. The new update has further improved its code generation skills, achieving an impressive score of 82.2% on the Aider Polyglot test. This score outperforms similar benchmarks set by OpenAI, Anthropic, and DeepSeek.
User Response
Users can expect the new model to be more innovative and provide well-formatted responses. This enhancement is a direct response to user feedback on the previous 03-25 update, especially regarding the performance of Gemini 2.5 Pro beyond coding tasks. Google acknowledges that this version addresses the regressions from the 03-25 update, demonstrating its commitment to user satisfaction.
New Feature
In addition, the latest update introduces configurable thinking budgets for developers, a feature initially introduced in the general-purpose Gemini 2.5 Flash model. This feature enables developers to customize the amount of time a Gemini thinking model allocates to a request within specified token limits. The objective is to optimize costs for developers integrating the model into their applications.
Benchmark Success
The updated Gemini 2.5 Pro model continues to excel in AI benchmarks, leading both the LMArena and WebDevArena leaderboards. Google also emphasizes that this version offers a more cost-effective per token operation compared to other comparable thinking models. The enhanced preview of Gemini 2.5 Pro is positioned as "enterprise-scale applications ready," facilitating seamless integration for developers.
Twitter Post
Google's latest Gemini 2.5 Pro update showcases improved performance in coding, reasoning, science, and math. It has demonstrated enhanced results across key benchmarks like AIDER Polyglot, GPQA, and HLE. The model has surged ahead in the @lmarena_ai rankings, gaining a 24pt Elo score increase from the previous version.
