previously Google Gemini 2.5 Flash and Gemini 2.5 Pro had already been announced, and after some time, with Google I/O 2025 arriving, these two models have become even more powerful. Gemini 2.5 Pro adds a brand-new Deep Think mode, while Gemini 2.5 Flash has improved across the board. Here’s a summary for everyone.
Gemini 2.5 Pro adds Deep Think reasoning mode
Gemini 2.5 Pro’s new DeepThink mode, according to Google, can push performance to the limit and deliver breakthrough results. This mode uses the latest cutting-edge research on thinking and reasoning, along with techniques like parallel processing, achieving top scores on the hardest math benchmarks currently available, including the USAMO 2025, and also leads competitors in LiveCodebench’s competition-level code evaluations.
At the same time, Gemini has always been a multimodal model, so it also performs excellently on multimodal tests like MMMU.
Here are the test scores shared by Google. Gemini 2.5 Pro’s DeepThink mode scored 49.4% on mathematics (USAMO 2025), which is about 15% higher than the 34.5% achieved without it. It also reached as high as 80.4% on LiveCodebenchv6, far ahead of OpenAI o4-mini’s 72.5%. On MMMU, it scored 84.0%, also higher than the other competitors:
However, Gemini 2.5 Pro DeepThink is still in its early stages at this point. Google has stated that it will spend more time on safety evaluations and have security experts conduct further review, so for now it is only available to trusted testers through the Gemini API, with a full rollout to follow later.
A stronger Gemini 2.5 Flash arrives
For the most efficient model, Gemini 2.5 Flash, Google has also continued to improve its capabilities. The updated 2.5 Flash shows improvements across all areas, including reasoning, coding, and long-context processing—key benchmark metrics where performance has clearly advanced.
On the LM Arena leaderboard, Gemini 2.5 Flash also ranks second, trailing only its own Gemini 2.5 Pro. However, the new Flash version won’t be released until early June, with Pro following shortly after.
Meanwhile, AI Studio, Vertex AI, and the Gemini App all have Gemini 2.5 Flash preview available to try out. If you’re interested, feel free to play around with it—the free tier works too.
Gemini 2.5 Flash and Gemini 2.5 Pro also gained next-generation voice capabilities.
In addition to the above, Google also announced that the voices of Gemini 2.5 Flash and Gemini 2.5 Pro have become more natural, allowing natural switching between “two different voices” within the same conversation, and the ability to mimic real conversation scenarios. Moreover, beyond ordinary intonation, they can express nuanced vocal emotions:
Supports over 24 languages, but it seems there’s no Chinese on the screen:
However, it is currently only available to the Gemini API:
Source: KOCPC Chinese