What you need to know
- Google’s beefed-up reasoning model, Gemini 2.5 Deep Think, is built to tackle advanced math, coding, and big-picture reasoning, and it is now live for AI Ultra subscribers.
- Deep Think supposedly works like OpenAI’s top models, piecing together smart solutions on the fly.
- It pulled an 84% on the MMMU test and crushed Olympiad questions, even if Google kept some numbers quiet.
Google announced today that it’s expanding access to an upgraded version of Gemini 2.5 Deep Think, now available to Google AI Ultra subscribers.
At Google I/O in May, the company unveiled major upgrades to its Gemini 2.5 lineup, including Deep Think, a new reasoning model for Gemini 2.5 Pro built to handle tougher challenges like higher-level math and coding.
Since the announcement, Google has been putting Deep Think through its paces. DeepMind CEO Demis Hassabis said back then that it’s built on “the latest cutting-edge research,” giving the model the ability to weigh multiple ideas at once before landing on a response.
Even after scoring big on the 2025 U.S. Math Olympiad, Google decided to run more safety checks and get expert input. So, it first gave Deep Think to a small group of trusted testers via the Gemini API to fine-tune things before going wide.
The advanced version is now rolling out to more users, giving them the tools to take on tough problems that require strategy, creativity, and deeper thinking, as per Google.
The company says the new model is a speedier spin on the one that snagged gold at the 2025 International Math Olympiad. Unlike the original, which could take hours to crack tough problems, this version is tuned for everyday tasks, but still holds its own with a Bronze-level score in internal IMO tests.
Strong benchmark results show Deep Think's potential
While the search giant did not share much about how Deep Think is built, it’s thought to work a lot like OpenAI’s o1-pro and o3-pro models, using some kind of engine to find and piece together the best solutions to a problem.
Gemini 2.5 Deep Think raises the bar for models that work without external tools. It topped the charts on LiveCodeBench V6, a competitive coding benchmark. Furthermore, it did well on Humanity’s Last Exam, a tough test covering advanced science and math.
Gemini 2.5 Deep Think adds even more value to the Google AI Ultra plan, which already includes early access to experimental video models, higher limits in tools like NotebookLM, a prototype AI agent manager, and a hefty 30TB of storage.
Google AI Ultra costs $250/month. If that’s too steep, the more budget-friendly AI Premium plan has been renamed Google AI Pro, and still sets you back $20 a year.
