Google Unveils Gemini 2.5 Pro Deep Think, Outperforming OpenAI’s O3 and O4 Models

Google I/O 2025: Major Enhancements to Gemini 2.5 Model Series Unveiled

At the recent Google I/O 2025 event, Google showcased significant advancements in its Gemini 2.5 model series, prominently featuring the introduction of the Gemini 2.5 Pro Deep Think mode. This innovation claims to surpass the performance of OpenAI’s most recent o3 and o4 model series in key AI benchmarks, marking a notable achievement in the field of artificial intelligence.

Enhancements in Reasoning Capabilities

Although there were no updates announced for the Gemini 2.5 Pro model, which underwent substantial upgrades recently, the launch of the Deep Think mode represents a paradigm shift in its reasoning capabilities. This new mode employs advanced research techniques to evaluate multiple hypotheses before generating a response, allowing for greater depth in understanding and problem-solving.

Benchmark Performance of 2.5 Pro Deep Think

Google revealed impressive benchmark results for the Gemini 2.5 Pro Deep Think mode, setting new standards of excellence:

49.4% on the 2025 USAMO math benchmarks.
80.4% on the LiveCodeBench competition-level coding benchmark.
84.0% on the MMMU multimodal reasoning benchmark.

These achievements establish the 2.5 Pro Deep Think mode as state-of-the-art (SOTA), exceeding the capabilities of OpenAI’s current offerings. Access to this advanced model will initially be restricted to trusted testers through the Gemini API.

Introduction of Gemini 2.5 Flash

In addition to the 2.5 Pro updates, Google introduced the Gemini 2.5 Flash model, designed for cost-effective usage. The Flash model has demonstrated performance improvements across all key benchmarks compared to its predecessor, and developers can preview it in Google AI Studio, Vertex AI for enterprise applications, and the Gemini app, with a broader rollout scheduled for June.

Enhancements to the Developer Experience

To enhance the Gemini platform for developers, Google also announced several key improvements:

A new Live API preview that supports multiple speakers, facilitating text-to-speech functionalities with dual voices through native audio output.
Integration of Model Context Protocol (MCP) definitions within the Gemini API for seamless interoperability with open-source tools.
The general availability of Gemini 2.5 Pro with thinking budgets, soon ready for stable production applications.
Project Mariner’s computing capabilities will be integrated into the Gemini API and Vertex AI.
Both 2.5 Pro and Flash models will now provide thought summaries through the Gemini API and Vertex AI.

For more in-depth information about these developments, you can view the announcement here.