Google’s Gemini 2.5 Pro Update Fixes Previous Model Issues
▼ Summary
– Google is releasing a major update to its Gemini 2.5 Pro AI model to fix issues from an earlier version, aiming for a stable public release.
– The new Gemini 2.5 Pro improves code generation, scoring 82.2% in the Aider Polyglot test, outperforming competitors like OpenAI and Anthropic.
– While Gemini 2.5 Flash has exited preview, the Pro version has faced criticism for performance outside coding tasks since the 03-25 update.
– Google has addressed feedback, with the new version improving creativity and response formatting to close gaps from previous regressions.
– The 06-05 release introduces configurable thinking budgets for developers and is expected to become a long-term stable release, dropping the “Preview” label soon.
Google’s latest Gemini 2.5 Pro update delivers significant improvements, addressing key issues from previous versions while enhancing performance across multiple areas. The tech giant continues refining its AI offerings, with this release specifically targeting stability and functionality gaps that emerged after earlier updates.
The new iteration follows the I/O Edition (05-06) which primarily focused on coding enhancements. Early benchmarks show impressive results, with the updated model achieving an 82.2% score on the Aider Polyglot coding test – outperforming competitors including OpenAI, Anthropic, and DeepSeek by notable margins. This positions Google’s solution as a strong contender for developers seeking advanced AI coding assistance.
While the lighter Gemini 2.5 Flash model has already exited preview status, the Pro version has faced delays in reaching similar maturity. Performance concerns surfaced following the 03-25 update, particularly regarding non-coding applications. Google’s engineering team, led by Logan Kilpatrick, acknowledges these challenges and confirms the current release specifically addresses previous shortcomings.
Key improvements include enhanced creative output and better response formatting, areas where users reported regression after earlier updates. The development team has implemented configurable thinking budgets, giving developers greater control over the model’s processing parameters. This flexibility represents a meaningful step forward for professional users requiring precise AI behavior.
Google anticipates this version (06-05) will serve as the long-term stable release, finally allowing the removal of the “Preview” label from consumer-facing applications. The rollout to public interfaces is expected within weeks, marking an important milestone in Gemini’s evolution from experimental technology to production-ready AI solution.
(Source: Ars Technica)
