Google has announced an important update to its advanced language model, Gemini 2.5 Pro, a step aimed at significantly enhancing its capabilities, particularly in the realm of coding and addressing feedback received on previous versions.
The new release comes as an improved preview version named Gemini 2.5 Pro Preview Thinking 06-05, which the company described as its smartest yet.
This version is expected to become the stable and generally available release for all users and enterprise applications within the next two weeks.
Performance Improvements and Addressing Previous Challenges
Tulsee Doshi, Google's Senior Product Management Director, stated that the updated Gemini 2.5 Pro reflects a significant improvement in performance.
The model showed a jump of 24 points on the Elo scale on the LMArena platform, maintaining its lead with a score of 1470 points.
It also achieved a gain of 35 points on the WebDevArena platform, reaching 1443 points.
These platforms rely on direct comparisons of outputs from different models by users, indicating a clear preference for the new model's outputs.
In this context, the model continues to excel in coding, leading challenging benchmark tests like Aider Polyglot, where it achieved an 82.2% score, outperforming competing models from companies such as OpenAI and Anthropic.
Additionally, the model demonstrated high-level performance in complex tests assessing math, science, knowledge, and reasoning capabilities, such as GPQA and Humanity's Last Exam (HLE).
This update came in response to feedback from users and developers regarding the previous version of Gemini 2.5 Pro.
According to Logan Kilpatrick from Google, the team worked to "close the loop on regressions" observed in the model's performance since the March 2025 (03-25) update, especially in non-coding tasks. Consequently, users can now expect more creative outputs with better response formatting.
Notable Accuracy and Enhanced Reasoning Capabilities
Testers noted that the new Gemini 2.5 Pro version puts more effort into delivering outputs with consistent formatting, such as using headings, bullet points, and bold text, elements that Google previously indicated were important based on user testing.
A striking example of the model's improved reasoning capabilities is its accurate and decisive answer to an old test question: "Would the color have been called 'magenta' if the town of Magenta had not existed?" The new model answered with a firm "No," which some other models failed to do or provided ambiguous answers.
Features for Developers and Wider Availability
For developers, the new release offers configurable "thinking budgets" within the Gemini API, available via Google AI Studio and Vertex AI. This feature gives developers greater control over cost and response time when building their applications.
The Google team expects this 05-06 version to become a "long-term stable release."
With its rollout to general users in the Gemini app and web interface over the coming weeks, the model is supposed to shed its "preview" tag.
The updated version of Gemini 2.5 Pro has already begun rolling out in the Gemini app and is also available to developers interested in experiencing it before general launch via Google AI Studio and Vertex AI platforms.
These updates represent part of Google's ongoing efforts to refine its intelligent models and provide more powerful and effective tools for users and developers alike.
