Google has presented Gemini 2.5 Pro, the first in its Gemini 2.5 family. This multimodal reasoning model surpasses the competitors of Openai, Anthrope and Deepseek in key reference points related to coding, mathematics and science.
What are reasoning AI models?
AIS reasoning is designed to “think before speaking.” They evaluate the context, the details of processes methodically and verify the answers to guarantee logical precision, although these capacities require more computer power and higher operating costs.
Openai launched the first reasoning model last September with O1, a notable deviation from the GPT series, which focused greatly on the generation of languages. Since then, the main players in AI's career have responded: Deepseek with R1, anthropic with Claude Sonnet 3.7 and Xai's with Grok 3.
Evolving beyond 'flash thought'
Google previously launched its first reasoning AI model, Gemini 2.0 Flash Thinking, in December. Marketing for his agent capabilities, Flash Thinking was recently updated to allow the loading of files and the largest indications; However, with the introduction of Gemini 2.5 Pro, Google seems to be withdrawing the label of “thought” completely.
According to Google's announcement on Gemini 2.5, this is because reasoning capabilities will now be integrated natively into all future models. This change marks a movement towards an architecture of the most unified, instead of separating the characteristics of “thinking” as an independent brand.
The new experimental model combines “a significantly improved base model” with “after training.” Google promotes its performance at the top of the LMarena classification table, which classifies the main large language models in several tasks.
Download: How to use Ia in Techrepublic Premium business
Reference leader in science, mathematics and code
Gemini 2.5 Pro stands out at academic reasoning points, scoring 86.7% in Aime 2025 (mathematics) and 84.0% at the reference point GPQA Diamond (Sciences). In the last examination of humanity, a broad test with thousands of questions in mathematics, sciences and humanities, the model leads with a score of 18.8%.
In particular, these results were achieved without the use of expensive testing techniques, which allow models such as O1 and R1 to continue learning during the evaluation.
At software development points, Gemini 2.5 Pro performance is mixed. He obtained 68.6% at Polyglot Aider's reference point for coding edition, surpassing most top -level models. However, it obtained 63.8% in the Swee Banco Verified, placing the second in the sonnet of Claude 3.7 in broader programming tasks.
Despite this, Google says that Gemini 2.5 Pro “stands out in the creation of web applications and visually convincing agent code applications”, as evidenced by its ability to create a video game from a single message.
The model admits a context window of one million tokens, which means that it can process the equivalent of a notice of 750,000 words, or the first six books of Harry Potter. Google plans to increase this threshold to two million tokens in due time.
Gemini 2.5 Pro is currently available through the Gemini Advanced application, which requires a subscription of $ 20 per month, and for developers and companies through Google Ai Studio. In the coming weeks, Gemini 2.5 Pro will be available in VERTEX AI, the Google Automatic Learning Platform for developers and price details will also be introduced for different fee limits.
