Google rolled Gemini Deep Think AI, a logic model that tests several ideas in parallel

Google Deepmind Gemini is rolling 2.5 deep think, which the company says, is the most advanced AI Reasoning model, which is able to answer questions by discovering and considering many ideas simultaneously and then to choose the best answer using those outputs.

Google’s $ 250-month ultra subscription subscribers will get access to Gemini 2.5 deep think in the Gemini app starting on Friday.

The first time in May/O was unveiled in May/O 2025, Gemini 2.5 Deep Think is the first publicly available multi-agent model of Google. These systems spon many agents to deal with a question in parallel, a process that uses much more computational resources than a single agent, but results in better answers.

Google used the variation of Gemini 2.5 deep think to score a gold medal in this year’s International Mathematics Olympiad (IMO).

With Gemini 2.5 Deep Think, the company says it is releasing the model used in the IMO to a select group of mathematicians and academics. Google says that this AI model “argues a few hours,” like most consumer-supporting AI models rather than second or minute. The company hopes the IMO model will increase research efforts, and its objective is to obtain a reaction to improve the multi-agent system for matters of educational use.

Google notes that Gemini 2.5 is a significant improvement on whatever has been declared in the 2.5 deep think model I/O. The company has also claimed to have developed “novel reinforcement learning techniques” to encourage Gemini 2.5 deep think to better use its argument paths.

In a blog post shared with Techcrunch, Google said, “Deep think can help people deal with problems that require creativity, strategic plan and need to improve step-by-step.”

The company says Gemini 2.5 deep thinks the final examination of humanity (HLE) achieves state-of-the-art performance-a challenging test measuring AI’s ability to answer thousands of crowds in humanities and science. Google claims that its model scored 34.8%on Hle (without equipment) compared to Grok 4 of XAI, which scored 25.4%and OPENAI’s O3, scoring 20.3%.

Google also states that Genini 2.5 Deep Think OPEAI, AI model from XAI, and anthropic, is a challenging test of competitive coding functions on Livecodebench6. Google’s model scored 87.6%, while Groke 4 scored 79%and Openai’s O3 scored 72%.

Gemini 2.5 deep thinkly works with equipment such as code execution and Google search, and the company says it is capable of producing “much longer response” compared to traditional AI models.

In Google’s test, the model produced more detailed and aesthetic web development works than other AI models. The company claims that the model can assist researchers and “potentially accelerate the passage to search.”

It seems that many major AI labs are changing around the multi-agent approach.

Elon Musk’s XAI recently released a multi-agent system of its own, Groke 4 Heavy, which states that it was capable of achieving the major performance of the industry on many benchmarks. Openai researcher Noam Brown said podcast The company used to win a gold medal in this year’s International Mathematics Olympiad (IMO) was also a multi-agent system. During this time, Anthropropic research agentWhich produces a fully research brief, is also operated by a multi-agent system.

Despite the strong performance, it seems that the multi-agent system is also expensive than the traditional AI model. This means that tech companies can place these systems behind their most expensive membership plans, which XAI and now Google have chosen to do it.

In the coming weeks, Google says that it is planning to share Gemini 2.5 deep think with a selected group of examiners through Gemini API. The company says that it wants to understand better how developers and enterprises can use their multi-agent system.