Google is rolling out “Deep Think,” the first public variant of its Gemini 2.5 family that adds extra “thinking time” and parallel reasoning to the model. From August 1, it appears in the Gemini mobile app, but only for Google AI Ultra subscribers, while a small group of mathematicians gets API access to the full Olympiad-grade build for research feedback. The subscription tier costs $249.99 per month after a three-month introductory price and includes a cap on daily Deep Think prompts.
Google appears to be quietly testing an internal variant of Gemini Deep Think, labelled “IMO” with a gold medal badge, likely a reference to its performance on the International Mathematical Olympiad benchmark, where the model achieved top-tier scores.
Gemini Deep Think IMO 👀
It is one of the first models which I am testing extensively b/c it is very fun to play with.
“Cyberpunk nuclear reactor control interface” https://t.co/X737eWXjUC pic.twitter.com/y5zHfZYm6Y
— TestingCatalog News 🗞 (@testingcatalog) August 1, 2025
Under the hood, the model keeps the sparse mixture-of-experts architecture of Gemini 2.5 Pro yet allocates longer inference budgets, letting it generate and critique many candidate solutions before responding. Internal papers list context windows above one million tokens, multimodal inputs, and tool use; industry analysts put total capacity near 1.5 trillion parameters, though Google hasn’t confirmed a figure. Benchmarks show double-digit gains over Gemini 2.5 Pro on LiveCodeBench V6 and Humanity’s Last Exam, and the “bronze” consumer build still reaches International Mathematical Olympiad bronze-level scores while running far faster than the gold competition model.
Had early access to Gemini with Deep Think. Very good model, big gains over standard Gemini 2.5 Pro for a lot of problems.
Here is the first attempt at the starship control panel prompt I try with every model. First time I have seen a model make a 3D interface in response. https://t.co/8iW2Pn6Xpu pic.twitter.com/bLFF2IcOP3
— Ethan Mollick (@emollick) August 1, 2025
Early users report material improvements: mathematicians say it helps explore conjectures; power users such as Wharton professor Ethan Mollick highlight first-time success on complex creative coding tests, noting the model produced a functional 3-D starship control panel from a single prompt.
For researchers, scientists, and academics tackling hard problems: Gemini 2.5 Deep Think is here. 🤯
It doesn’t just answer, it brainstorms using parallel thinking and reinforcement learning techniques. We put it into the hands of mathematicians who explored what it can do ↓ pic.twitter.com/lsBq3ICvuM
— Google DeepMind (@GoogleDeepMind) August 1, 2025
For Google, Deep Think reinforces its strategy of tiered releases, Flash for speed, Pro for balance, Deep Think for heavy reasoning, aimed at keeping premium subscribers inside its AI ecosystem. The company is also using the launch to gather safety data on extended-reasoning systems before wider API distribution, a step that aligns with its broader push to pair frontier research with tightly managed roll-outs in consumer products.
