Gemini’s next-generation model is here! Earlier, Google announced the launch of the new Gemini 2.5 model. Like Gemini 2.0 Flash Thinking, Gemini 2.5 is also a thinking model that reasons through problems on its own before answering, improving performance and accuracy. Google also shared benchmark comparisons against other AI models, and Gemini 2.5 delivers exceptionally strong results.
Google’s new-generation Gemini 2.5 model officially unveiled.
According to Google, the new generation Gemini 2.5 shows notable improvements across tasks, including coding, support for longer context (now 1 million tokens, with plans to upgrade to 2 million), and more. On the model side, only Gemini 2.5 Pro is available for now.
On the LMArena leaderboard (which evaluates model capabilities based on human preference), Gemini 2.5 Pro took first place as soon as it launched, leading Grok-3-Preview and GPT-4.5-Preview with a score of 1443, meaning that in terms of language style and expression quality, it is currently the strongest model:
Gemini 2.5 Pro’s coding capabilities have improved significantly, now covering the creation of visual web applications, agentic coding applications, and code conversion and editing tasks.
With a score of 63.8% in the SWE-Bench Verified agentic coding test, it surpassed OpenAI o3-mini-High’s 49.3% and DeepSeek R1’s 49.2%, but still slightly lagged behind Claude 3.7 Sonnet’s 70.3%.
The Aider polyglot cross-language coding test also achieves 74.0%, far ahead of o3-mini-High, Claude 3.7 Sonnet, and DeepSeek R1.
Humanity’s Last Exam, a comprehensive intelligence test, is no exception, scoring higher than all competitors at 18.8%.
Basically, Gemini 2.5 Pro scored the highest in almost all tests:
Google also shared a demonstration video showing how to use a single line of prompt to get Gemini 2.5 Pro to write a game:
Gemini 2.5 Pro also supports multimodal capabilities, handling text, audio, images, video, entire codebases, and more.
Developers and enterprise users can now experience Gemini 2.5 Pro in Google AI Studio starting today, while Gemini Advanced paid subscribers can also find this version in the model menu on desktop and mobile. Vertex AI support will be rolling out within a few weeks.
Free users can also Google AI Studio To experience it, once you enter the website, you can switch to Gemini 2.5 Pro Experimental from the Model menu on the right:
Source:Google
Source: KOCPC Chinese