Credit: Ryan Whitwam

It seems like hardly a day goes by anymore without a new version of Google's Gemini AI landing, and sure enough, Google is rolling out a major update to its most powerful 2.5 Pro model. This release is aimed at fixing some problems that cropped up in an earlier Gemini Pro update, and the word is, this version will become a stable release that comes to the Gemini app for everyone to use.

The previous Gemini 2.5 Pro release, known as the I/O Edition, or simply 05-06, was focused on coding upgrades. Google claims the new version is even better at generating code, with a new high score of 82.2 percent in the Aider Polyglot test. That beats the best from OpenAI, Anthropic, and DeepSeek by a comfortable margin.

While the general-purpose Gemini 2.5 Flash has left preview, the Pro version is lagging behind. In fact, the last several updates have attracted some valid criticism of 2.5 Pro's performance outside of coding tasks since the big 03-25 update. Google's Logan Kilpatrick says the team has taken that feedback to heart and that the new model "closes [the] gap on 03-25 regressions." For example, users will supposedly see more creativity with better formatting of responses.

Kilpatrick also notes that the 06-05 release now supports configurable thinking budgets for developers, and the team expects this build to become a "long term stable release." So, Gemini 2.5 Pro should finally drop its "Preview" disclaimer when this version rolls out to the consumer-facing app and web interface in the coming weeks.

Credit: Google

Google is very interested in the vibes of its models, so it often leans on the ratings from LMArena and WebDevArena. These platforms have users compare model outputs side by side to rate the outputs before revealing what they are. Gemini 2.5 debuted earlier this year with a top score on LMArena, suggesting people like its outputs more than every other large language model.

Again, Google is crowing about being at the top of the heap, with a 24-point Elo score increase on LMArena and a jump of 35 Elo points on WebDevArena. Google was already in the top slot on both leaderboards, so it has just increased its lead with the latest Gemini Pro.

In our testing, the new model makes a greater effort to use coherent formatting in the output. Google's AI team has noted on several occasions that things like headers, bullets, and bold text are surprisingly beneficial based on the company's user testing.

We've also noticed Gemini 2.5 Pro finally answers one of our long-running test questions accurately and with authority. When asked, "Would the color be called 'magenta' if the town of Magenta didn't exist?" the new Gemini 2.5 Pro correctly answers with an emphatic "no."

If you want to tinker with the next Gemini Pro before final release, maybe to ask it other obscure color questions, it's available today in Vertex AI and AI Studio. It's also coming to the Gemini app (in preview) starting today.