Gemini 2.5 Pro tops AI coding tests, surpasses ChatGPT and Claude

Google has released an updated version of its Gemini 2.5 Pro model, addressing issues found in earlier updates.

Unlike the I/O Edition, which focuses mostly on coding, the new version improves performance more broadly and is expected to become a stable release in both the Gemini app and web interface.

The company claims the updated model performs significantly better in code generation, topping the Aider Polyglot test with a score of 82.2 percent—surpassing offerings from OpenAI, Anthropic and DeepSeek.

Beyond coding, the model aims to close previous performance gaps introduced with the March 25th update, especially in creativity and response formatting.

Developers can now fine-tune the model’s ‘thinking budget’, while users should notice a more transparent output structure. These changes and consistent improvement in leaderboard ratings on LMArena and WebDevArena suggest that Google is extending its lead in the AI race.

Google continues to rely on blind testing to judge how people feel about its models, and the new Gemini Pro seems to resonate well. In fact, it now answers even quirky test questions with more clarity and confidence—something that had been lacking in earlier versions.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!