List your company

Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

About

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Key insights

What problem does Gemini 2.5 address?

The paper introduces the Gemini 2.X family, addressing the need for models that combine advanced reasoning, multimodal understanding, and long context for complex agentic tasks. It uses sparse mixture-of-experts transformers with a "thinking" capability for inference-time computation.

How does it differ from prior work?

Gemini 2.5 Pro improves over Gemini 1.5 Pro, raising AIME 2025 from 17.5% to 88.0% and LiveCodeBench from 29.7% to 74.2%. Evaluations span coding, math, factuality, and long-context benchmarks.

What are the practical applications?

The report describes use in Google products including AI Overviews, Gemini Deep Research and NotebookLM. It separately demonstrates agent workflows such as playing Pokémon and turning video into interactive applications; those demonstrations are not evidence of deployment in those tasks.

What limitations and future work exist?

The authors note that as context grows beyond 100k tokens, agents favor repeating past actions over synthesizing novel plans. They also state that developing sufficiently challenging evaluation benchmarks has struggled to keep pace with model improvements.

Related benchmarks