Following updates to the models that are available for all users, Google today announced Gemini 2.5 Pro (experimental) for Advanced subscribers and developers. Like before, Google is doing another mid-year/model update.
Notably, all models in the Gemini 2.5 family, including futures ones, are “thinking models, capable of reasoning through their thoughts before responding, resulting in enhanced performance and improved accuracy.” Google is “building these thinking capabilities directly into all of [its] models” to allow them to “handle more complex problems and support even more capable, context-aware agents.”
Compared to 2.0 Flash Thinking, which was first revealed in December and got an update this month, Google is no longer explicitly attaching the “Thinking” label. Users can “Show thinking” in the Gemini app to see the train of thought.
In the field of AI, a system’s capacity for “reasoning” refers to more than just classification and prediction. It refers to its ability to analyze information, draw logical conclusions, incorporate context and nuance, and make informed decisions.
Gemini 2.5 features a “new level of performance by combining a significantly enhanced base model with improved post-training.”
Gemini 2.5 Pro (gemini-2.5-pro-exp-03-25 and codenamed “nebula”) is the first model in this family. Aimed at complex tasks, Google notes how it “tops the LMArena leaderboard — which measures human preferences — by a significant margin.” It also leads on math (AIME 2025), and science (GPQA diamond) benchmarks “without test-time techniques that increase cost, like majority voting.”
It also scores a state-of-the-art 18.8% across models without tool use on Humanity’s Last Exam, a dataset designed by hundreds of subject matter experts to capture the human frontier of knowledge and reasoning.
Top comment by Omega192
Dang, that was a quick .5 jump. I wonder if the recent batch of announcements was already planned or they pushed them out early to capitalize on OpenAI's last two underwhelming and overpriced releases (GPT-4.5 and o1-pro).
Either way, pretty impressive this model scores on par or better than the competition without relying on test-time compute. Real curious to hear more about how that was accomplished. Seems Gemini remains unmatched when it comes to long context, too.
I somehow missed that o3-mini and R1 weren't multimodal before now. Kinda seems like OpenAI split their multimodal and thinking efforts up while DeepMind carried on with a single model for both. Guess we'll see if o4/GPT-5 follows that lead.
There’s also a focus on advanced coding with a “big leap over 2.0” and “more improvements to come.”
- “2.5 Pro excels at creating visually compelling web apps and agentic code applications, along with code transformation and editing.”
- “On SWE-Bench Verified, the industry standard for agentic code evals, Gemini 2.5 Pro scores 63.8% with a custom agent setup.”
In addition to native multimodality, Gemini 2.5 Pro has a 1 million token context window with 2 million coming soon.
It can comprehend vast datasets and handle complex problems from different information sources, including text, audio, images, video and even entire code repositories.
Gemini 2.5 Pro (experimental) is rolling out first to Gemini Advanced and Google AI Studio, with Vertex AI following in the coming weeks.
We’ll also introduce pricing in the coming weeks, enabling people to use 2.5 Pro with higher rate limits for scaled production use.
In the Gemini app, it replaces last month’s 2.0 Pro (experimental) with access to apps (@Gmail, @YouTube, etc.) and file uploads.
FTC: We use income earning auto affiliate links. More.
Comments