Google Unveils Gemini 2.5 Deep Think: A New Benchmark in AI Language Models

Google has announced the launch of Gemini 2.5 Pro Deep Think, its most powerful and expensive language model to date. This new offering will be available to subscribers of the Gemini Ultra plan, priced at $250 per month, while API access is still being tested. Additionally, a specially fine-tuned version of Deep Think, which previously earned a gold medal at the 2025 International Mathematics Olympiad (IMO 2025), will be accessible to a select group of mathematicians.

The Deep Think mode is based on the Gemini 2.5 Pro model but features several unique attributes. Firstly, it boasts enhanced computational resources for reasoning, which is crucial for tasks that require prolonged problem-solving. Secondly, Deep Think utilizes multiple AI agents that work in parallel to explore various approaches to a given problem. Ultimately, a critique module either selects the optimal solution or synthesizes a final answer from several options.

Google compared the performance of Gemini 2.5 Deep Think to other versions of Gemini 2.5 while generating a voxel scene:

The company shared benchmarks from Deep Think’s performance without utilizing any external tools:

It’s worth noting that Google did not include results from GPT o3 Pro and Grok 4 Heavy in the comparisons, likely due to differing testing methodologies, as xAI and OpenAI provided their models with access to tools.

P.S. You can support me by subscribing to the channel «runaway neural network«, where I explore the creative aspects of AI.