Google starts rolling out its upgraded AI language model, Gemini 2.5 Pro, to select users: Check features, eligibility, and more

Google Gemini

Google has rolled out a preview version of its upgraded AI model, Gemini 2.5 Pro, to select developers and enterprise users. The model is accessible through the Gemini API via Google AI Studio, Vertex AI, and the Gemini app. General availability is expected in the coming weeks.

Gemini 2.5 Pro is available in a limited preview. Developers can begin using it immediately on Google AI Studio and Vertex AI. Google has also introduced thinking budgets, allowing users to manage inference cost and latency. The rollout has also started on the Gemini mobile app.

The model has posted significant gains on major benchmarks. It recorded a 24-point Elo increase on LMArena (now at 1470) and a 35-point increase on WebDevArena (now at 1443). These gains reflect better performance in tasks requiring long-term reasoning and software development.

Gemini 2.5 Pro leads or ranks highly across a range of industry-standard tests. According to Google’s benchmark chart:

• Science (GPQA diamond, single attempt): 86.4%

• Mathematics (AIME 2025, single attempt): 88.0%

• Code generation (LiveCodeBench): 69.0%

• Code editing (Aider Polyglot): 82.2%

• Visual reasoning (MMMU, single attempt): 82.0%

• Image understanding (Vibe-Eval): 67.2%

• Video understanding (VideoMMMU): 83.6%

• Agentic coding (SWE-bench Verified, single attempt): 59.6%

• Factuality (FACTS Grounding): 87.8%

• Long context (MRCR v2, 8-needle): 58.0%

• Multilingual performance (Global MMLU Lite): 89.2%

Google has improved response structure and formatting in the upgraded model. This addresses developer feedback from the earlier Gemini 2.5 Pro release. The model also offers stronger performance in creative and structured tasks, supported by tools like Aider for code editing.

Gemini 2.5 Pro integrates with Google’s enterprise platforms and supports parallel compute and majority voting in multiple-attempt modes. It is intended for high-scale deployments where accuracy and efficiency are critical.