Google just rolled out Gemini 2.5 Deep Think to subscribers of its Google AI Ultra service. This new model takes more time to work through complex problems before giving you an answer.
Google CEO Sundar Pichai announced the new version on X:
Deep Think first appeared during Google I/O in May 2025 and an experimental version was made available to selected partners and developers through early-access programs on the Vertex AI platform and Gemini API. Today’s consumer launch brings an improved model that incorporates months of feedback from early testers and new research breakthroughs.
We’re bringing a version of Deep Think that achieved gold-medal status at IMO to Ultra subscribers in the @Geminiapp (+ the official version is now in the hands of mathematicians).
— Sundar Pichai (@sundarpichai) August 1, 2025
Toggle it on when reasoning through complex scientific literature, tackling a coding problem that… pic.twitter.com/OyFSGsQSgJ
If you subscribe to Google AI Ultra, you can start using the new Deep Think version today through the Gemini app. You’ll see a toggle option for “Deep Think” when you select Gemini 2.5 Pro from the model dropdown menu. Keep in mind you get a limited number of prompts per day.
What makes Deep Think different
Most generative AI models respond instantly. For many, that’s the appeal.
Deep Think doesn’t. Instead, it uses what Google calls “parallel thinking techniques” to generate multiple ideas at once, compare them, and refine its approach before settling on the best solution.
Think of it like this: When you face a tough problem, you don’t just say the first answer that comes to mind. You consider different angles, weigh your options, and think through the consequences. That’s what Deep Think does.
If you want to ask “when was Google founded and by whom?” then Deep Think is not the best option as it’s not a complex question.
However, if you provide it with a research paper and ask for real-world applications the authors didn’t explore, it will analyze the work more thoroughly than standard models. While regular AI might offer quick, surface-level suggestions, Deep Think examines the research from multiple angles to identify practical uses the original researchers may have overlooked.
Of course, there are many other applications; the example above serves to drive the point that this advanced model is catered to problems that require advanced reasoning.
Just how good is it?
Google tested Deep Think on several benchmarks. On LiveCodeBench V6, which measures competitive coding ability, Deep Think scored 87.6%.
For Humanity’s Last Exam, a test that measures expertise across science, math, and other domains, Deep Think achieved 34.8%. While this beats other models shown in Google’s comparison, Grok 4’s “Heavy” version with Python and internet access has scored 44.4% on the same test. Regardless, it’s a huge feat for Google.
Deep Think’s strongest showing came in mathematics. On the IMO 2025 benchmark, it reached 60.7%, earning a bronze medal grade and significantly outperforming other models. It scored even higher on AIME 2025 at 99.2%.
You may be wondering, wait didn’t Google say one of their models earned a gold medal grade on the IMO? Yes, and the company said it will also be giving select mathematicians and academics access to the full version that competed in the International Mathematical Olympiad. This version takes hours to solve problems but did achieve gold-medal performance.
The consumer version works faster while still delivering bronze-level results on the same benchmark.
August, a big month for AI model releases
Google’s new version of Deep Think comes at an interesting time. OpenAI is expected to release its next big model, GPT-5, sometime this month August according to sources close to the company. This puts Google in a race against time to establish Deep Think before the AI landscape shifts again.
OpenAI CEO Sam Altman recently posted on X that GPT-5 would be released ‘soon,’ and appeared on Theo Von’s podcast, where he praised the abilities of the new model, which is expected to combine the attributes of both traditional models and so-called reasoning models such as o3, potentially making it a direct competitor to Deep Think’s parallel thinking approach.
The timing matters. Code references to a model called ‘gpt-5-reasoning-alpha-2025-07-13’ have been spotted on X, suggesting the final round of testing is underway.
