Google has released Gemini 2.5 Pro Experimental, touting it as their most advanced 'reasoning' AI model to date. Available through Google AI Studio and Gemini Advanced, this multimodal model supports text, image, audio, and video inputs, outputting text with a context window of up to 1 million tokens, expandable to 2 million. Gemini 2.5 Pro excels in coding, complex reasoning, and handling diverse inputs. It can call external functions, generate structured output, execute code, and utilise search, making it adept at multi-step tasks and API calls.
Benchmark tests reveal Gemini 2.5 Pro's strengths in mathematics, logic, and long-context tasks. It leads in benchmarks like AIME 2024 and GPQA Diamond. While it performs well in code generation, it doesn't dominate in all coding benchmarks. However, it stands out in long-context reading comprehension, achieving 91.5% accuracy with a 128,000 context length. Google aims to bring the best model to all Gemini users, gathering feedback to improve models over time. The model is available to all Gemini users with rate limits, while Gemini Advanced users have expanded access and a larger context window.