Google unveils "thinking model" Gemini 2.5 Pro — takes a commanding lead on LMArena
Google announced its new model, “Gemini 2.5 Pro.” A reasoning-focused model that improves accuracy by going through an internal “thinking” process before responding, it shot to the top of the AI model comparison site LMArena by a wide margin immediately after release, drawing significant attention.
Details
- Thinking model: Comes standard with a “thinking” capability that reasons through complex problems step by step before producing an answer
- Context window: Supports up to 1 million tokens (with an expansion to 2 million tokens announced as forthcoming). Can process text, audio, images, video, and entire code repositories together
- Benchmarks: Took a commanding lead on LMArena. Scored 18.8% without tool use on the challenging “Humanity’s Last Exam” benchmark
- Coding performance: Achieved 63.8% on SWE-Bench Verified using a custom agent setup, earning high praise from developers
- Availability: Made available immediately to Google AI Studio and Gemini Advanced users through the Gemini app. Rollout to Vertex AI was to follow, with pricing to be announced later
What happened next
Gemini 2.5 Pro was positioned as the flagship of Google’s “thinking model” lineup, receiving further updates at Google I/O that May, and became the core driver of reasoning improvements across the Gemini series. Its strong reputation in coding also led to broader integration with developer tools.