On Tuesday, Google introduced Gemini 2.5, a new family of AI reasoning models designed to pause and “think” before responding to questions.
To launch this new series, Google released Gemini 2.5 Pro Experimental, a multimodal and reasoning AI model claimed to be the company’s most sophisticated to date. Starting Tuesday, this model will be accessible on Google’s developer platform, Google AI Studio, and through the Gemini app for subscribers of the Gemini Advanced plan, which costs $20 per month.
Google has announced that all of its future AI models will incorporate reasoning capabilities.
Since OpenAI presented the first AI reasoning model, o1, in September 2024, the tech industry has been striving to match or surpass this model’s capabilities. Currently, companies like Anthropic, DeepSeek, Google, and xAI have developed AI reasoning models that utilize additional computing power and time to fact-check and logically process problems before providing answers.
These reasoning techniques have allowed AI models to achieve significant advancements in mathematical and coding tasks. The tech community views reasoning models as likely key components of AI agents, which are autonomous systems capable of performing tasks with minimal human intervention. However, these models come at a higher cost.
Google has previously experimented with AI reasoning models, having released a “thinking” version of Gemini in December. Gemini 2.5, however, is Google’s most determined effort to surpass OpenAI’s “o” series of models.
According to Google, Gemini 2.5 Pro surpasses its preceding frontier AI models and some of the leading competing AI models in various benchmarks. The model is specifically designed to excel in creating visually appealing web apps and sophisticated coding applications.
In the Aider Polyglot evaluation, which measures code editing, Gemini 2.5 Pro achieved a score of 68.6%, surpassing the leading AI models from OpenAI, Anthropic, and Chinese AI lab DeepSeek.
However, in the SWE-bench Verified test, which evaluates software development abilities, Gemini 2.5 Pro scored 63.8%. This performance exceeded OpenAI’s o3-mini and DeepSeek’s R1 but was below Anthropic’s Claude 3.7 Sonnet, which achieved 70.3%.
In the Humanity’s Last Exam, a multimodal test featuring thousands of crowdsourced questions in mathematics, humanities, and natural sciences, Gemini 2.5 Pro scored 18.8%, outperforming most competitor flagship models.
Initially, Google’s Gemini 2.5 Pro will feature a context window of 1 million tokens, allowing the AI model to process approximately 750,000 words simultaneously, a scope larger than the entire “Lord of The Rings” book series. Soon, the model will support input lengths up to 2 million tokens.
Google has not yet disclosed the API pricing for Gemini 2.5 Pro, with additional details expected to be released in the coming weeks.
