Google is sharpening the reasoning abilities of its most capable Gemini AI models, with a new mode designed to make Gemini 2.5 Pro work through harder questions more carefully before it answers.
The company announced Deep Think on Tuesday at Google I/O 2025. It is described as an enhanced reasoning mode for Gemini 2.5 Pro, and Google says it can improve performance on certain benchmarks by letting the model consider multiple answers before producing a response.
What Deep Think changes in Gemini 2.5 Pro
Deep Think is not being presented as a separate model. It is a mode for Google’s flagship Gemini 2.5 Pro model, built to push more reasoning into the process that happens before an answer appears.
That distinction matters. The core promise is not simply faster output or a larger response. It is that Gemini 2.5 Pro can evaluate more than one possible path when faced with a question, then use that extra internal work to improve the quality of the final answer.
Demis Hassabis, head of Google DeepMind, described the feature during a press briefing as a way to push model performance to its limits. He said Deep Think uses Google’s latest research in thinking and reasoning, including parallel techniques.
Google did not give a detailed technical explanation of how Deep Think works. The source article notes that it could be similar to OpenAI’s o1-pro and upcoming o3-pro models, which likely use an engine to search for and synthesize the best solution to a given problem.
Why benchmarks are central to the announcement
Google framed Deep Think around measurable improvements. According to the company, Gemini 2.5 Pro with Deep Think topped LiveCodeBench, described in the source as a challenging coding evaluation.
Google also said Gemini 2.5 Pro Deep Think beat OpenAI’s o3 on MMMU, a test involving skills such as perception and reasoning. Those claims position Deep Think as a feature aimed at the kinds of problems where a model has to do more than retrieve or summarize information.
The benchmark focus also shows where Google sees the value of enhanced reasoning. Coding, perception, multimodal understanding, and structured reasoning are all areas where a model can fail if it rushes to a single answer too quickly.
For users, the practical idea is straightforward: a reasoning mode can be useful when the question has competing possible solutions, when a coding task has multiple implementation paths, or when the model needs to combine different kinds of information before answering.
Access is limited while safety work continues
Deep Think is not broadly available yet. Google said it is available to trusted testers through the Gemini API as of this week.
The company is taking additional time for safety evaluations before a wider rollout. That means developers and other users outside the trusted tester group will have to wait before they can use Deep Think more broadly.
This staged release is an important part of the announcement. Google is pairing a performance claim with a slower access plan, suggesting that stronger reasoning capabilities are being handled cautiously before they reach a wider audience.
For now, the key facts are limited but clear:
- Deep Think is an enhanced reasoning mode for Gemini 2.5 Pro.
- It lets the model consider multiple answers before responding.
- Google says it improved results on LiveCodeBench and MMMU.
- It is available to trusted testers through the Gemini API as of this week.
- Google is conducting safety evaluations before broader release.
Gemini 2.5 Flash also gets a performance update
Deep Think was not the only Gemini update announced. Google also introduced an update to Gemini 2.5 Flash, its budget-oriented model.
The new Gemini 2.5 Flash is meant to perform better on tasks involving coding, multimodality, reasoning, and long context. Google also says it is more efficient than the version it replaces.
The updated 2.5 Flash is available for preview in Google’s AI Studio and Vertex AI platforms, as well as in the company’s Gemini apps. Google says the improved Gemini 2.5 Flash will become generally available for developers sometime in June.
This gives Google two different Gemini updates aimed at different needs. Gemini 2.5 Pro with Deep Think is positioned around higher-end reasoning performance, while Gemini 2.5 Flash is being improved as a more efficient option for a broader set of tasks.
Gemini Diffusion points to speed
Google also introduced Gemini Diffusion, a model the company claims is very fast. According to Google, it can deliver output 4-5 times quicker than comparable models and rival the performance of models twice its size.
Gemini Diffusion is available beginning today to trusted testers. The source does not provide further details about how the model works, but Google’s framing centers on speed and performance relative to model size.
Taken together, the announcements show Google pushing Gemini in several directions at once: deeper reasoning for Gemini 2.5 Pro, broader efficiency and capability gains for Gemini 2.5 Flash, and faster output through Gemini Diffusion.
The most closely watched piece may be Deep Think, because it goes directly at the question of how AI models handle complex problems. Google has not yet explained the full mechanics, but it has made the product direction clear: Gemini 2.5 Pro is being upgraded to think through more possibilities before it speaks.