Google is a Gemini Deepton Out-Out Eye, a reasoning model that tests multiple ideas in parallel

Google DeepMind is being deployed Gemini 2.5 Deep Thinkthe company says it is the most advanced AI inference model, allowing you to answer questions by researching and examining multiple ideas simultaneously and using their output to select the best answer.

Google $250/month subscriber Ultra Subscriptions will be able to access Gemini 2.5 Deep Think through the Gemini app starting Friday.

First unveiled in May at Google I/O 2025, the Gemini 2.5 Deep Think is Google’s first published multi-agent model. These systems work on problems to work on parallel questions, a process that generates AI multiple agents and uses much more computational resources than a single agent, but tends to provide better answers.

Google used a variation of Gemini 2.5 Deep Think Win a gold medal This year’s International Mathematics Olympiad (IMO).

In addition to Gemini 2.5 Deep Think, the company says it is releasing models used in IMO to select groups of mathematicians and scholars. Google says this AI model “takes hours to infer,” rather than seconds or minutes like most consumer AI models. The company hopes that the IMO model aims to enhance research efforts and get feedback on how to improve multi-agent systems for academic use cases.

Google points out that the Gemini 2.5 Deep Think model is a significant improvement over what it announced in I/O. The company also claims it has developed a “new reinforcement learning technology” to encourage Deep in Gemini 2.5.

“Deep thinking helps people tackle issues that require creativity, strategic planning and improvement in stages,” Google said in a blog post that it shares with TechCrunch.

The company says the Gemini 2.5 Deep Think is achieving cutting-edge performance in the final test of humanity (HLE). This is a challenging test that measures your ability to answer thousands of crowdsourced questions to thousands of crowdsourced questions. Google claims that the model scored 34.8% (no tool) in HLE compared to Xai’s Grok 4, which scored 25.4% and Openai’s O3, which scored 20.3%.

Google also states that the Gemini 2.5 Deep Think is superior to the Openai, Xai and human AI models in LiveCodebench6, a challenging test of competitive coding tasks. Google’s model won 87.6%, while Grok 4 won 79%, while Openai’s O3 won 72%.

Gemini 2.5 Deep Think works automatically with tools like code execution and Google search. The company says it can create “a much longer response” than traditional AI models.

In Google’s testing, the model created a more detailed and aesthetically pleasing web development task compared to other AI models. The company claims that the model will support researchers and “potentially accelerate the path to discovery.”

It appears that several major AI labs are converging around a multi-agent approach.

Elon Musk’s Xai recently released its own multi-agent system. Groke 4 Heavyit says it was able to achieve industry-leading performance in several benchmarks. Noam Brown, a researcher at Openai, Podcasts The unreleased AI model the company used to win gold medals at this year’s International Mathematics Olympiad (IMO) was also a multi-agent system. meanwhile, Human Research AgentIt also features a multi-agent system that generates a thorough research brief.

Despite its strong performance, multi-agent systems seem to be even more expensive to service than traditional AI models. This means that tech companies may store these systems behind the most expensive subscription plans.

Over the next few weeks, Google says it will be sharing Gemini 2.5 Deep Think with groups of selected groups via the Gemini API. The company wants to better understand how developers and businesses use multi-agent systems.