News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
imgwavvidGemini 2.5 Pro

Gemini 2.5 Pro

Google DeepMind's thinking model for complex reasoning, code and long-context work

Gemini 2.5 Pro is a multimodal "thinking" model from Google DeepMind, first shown as an experimental release in March 2025. It reasons through its work before answering and handles text, audio, images, video and whole codebases, with a 1 million token context window at launch.

Share on WhatsApp

  • Launched March 2025 as Gemini 2.5 Pro Experimental
  • 1 million token context window at launch, 2 million announced as coming
  • Reads text, audio, images, video and whole code repositories
  • Google reports 63.8% on SWE-Bench Verified with a custom agent setup
  • Google reports 18.8% on Humanity's Last Exam without tool use

What it is

Gemini 2.5 Pro is Google DeepMind's model for hard problems, built to think through its reasoning before it replies. The company describes it as its most intelligent model at the time of the March 2025 launch. It is multimodal and long-context: it can read text, audio, images, video and entire code repositories, with a window of 1 million tokens at release and 2 million promised later. Google positions it for complex reasoning, maths, science and coding work, including building web apps and agentic code. It arrived first as an experimental version, then moved into wider product use.

How it came about

Google DeepMind introduced the Gemini 2.5 line in March 2025, opening with an experimental build of 2.5 Pro. The company had earlier shipped a thinking model, Gemini 2.0 Flash Thinking, and says 2.5 combines a much stronger base model with better post-training. The launch blog was updated on 26 March with new multi-round coreference resolution evaluations. A technical report on the Gemini 2.5 family was posted to arXiv in July 2025 and revised through December 2025.

Inside the model

Google describes 2.5 as a thinking model: it works through its reasoning internally before answering, which the company says lifts accuracy. The approach builds on reinforcement learning and chain-of-thought prompting used in earlier work. The model keeps the Gemini line's native multimodality and long context, taking in text, audio, images, video and full codebases through a 1 million token window at launch. Google says it pairs a significantly improved base model with stronger post-training, and plans to build this reasoning behaviour into all its models.

What it can do

Google reports leading results on common coding, maths and science benchmarks without extra test-time tricks such as majority voting, naming GPQA and AIME 2025. It cites 18.8% on Humanity's Last Exam without tool use, and 63.8% on SWE-Bench Verified with a custom agent setup. The company highlights web app creation, agentic code, and code transformation and editing, and showed it building a video game from a one-line prompt. It also topped the LMArena human-preference leaderboard at launch.

How to use it

At launch, Gemini 2.5 Pro Experimental was available in Google AI Studio and to Gemini Advanced users in the Gemini app, with Vertex AI to follow. Google said pricing would arrive in the following weeks, allowing higher rate limits for production use.

Reception and debate

Google says the model debuted at number one on LMArena, a leaderboard of human preferences, by a significant margin, which it reads as evidence of strong capability and response style. The company frames the release as experimental and asks users for feedback to guide further improvements.

Why it matters

Google says reasoning of this kind is being built into all of its models so they can take on harder problems and support more context-aware agents. 2.5 Pro is presented as a large step up from Gemini 2.0 in coding, with further gains planned.

Questions people ask about Gemini 2.5 Pro

What makes Gemini 2.5 Pro a "thinking" model?

It reasons through its internal chain of thought before giving an answer. Google DeepMind says this comes from combining a stronger base model with improved post-training, building on reinforcement learning and chain-of-thought prompting used in its earlier Gemini 2.0 Flash Thinking model.

What can it take as input?

It is natively multimodal, handling text, audio, images, video and even whole code repositories. At launch it had a 1 million token context window, with Google saying a 2 million token version was coming.

How do I use it?

At launch it was in Google AI Studio and selectable by Gemini Advanced users in the Gemini app on desktop and mobile. Google said Vertex AI support and pricing for higher-rate production use would follow in the coming weeks.

How good is it at coding?

Google reports 63.8% on SWE-Bench Verified using a custom agent setup, and highlights web app creation, agentic code applications, and code transformation and editing. It showed the model producing a working video game from a single-line prompt.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Gemini 2.5 Pro yet, so the facts on this page come from the sources listed above: blog.google/technology/google-deepmind/gemini-model-thinking-updates-m, arxiv.org/abs/2507.06261. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.