Google DeepMind Releases Gemini 2.5 Pro With Enhanced Reasoning

What Happened

Google DeepMind released Gemini 2.5 Pro on March 25, 2025, making the model available to developers through Google AI Studio and the Gemini API. The release marks the first public availability of a model from the Gemini 2.5 generation, which Google describes as its most capable series to date.

Gemini 2.5 Pro supports a context window of one million tokens, matching the capacity introduced in earlier Gemini 1.5 models while pairing it with what Google characterises as substantially improved reasoning performance.

Background

Google DeepMind is the consolidated AI research and product division of Alphabet Inc., formed in 2023 through the merger of Google Brain and DeepMind. The division competes directly with OpenAI, Anthropic, and Meta in the large language model market.

The Gemini model family launched in December 2023 as the successor to Google's PaLM 2 series. Gemini 1.0 was followed by Gemini 1.5 in February 2024, which introduced the one-million-token context window as a key differentiator. Gemini 2.0 Flash followed in late 2024, with Google positioning it as a faster, more efficient option for developers.

Gemini 2.5 Pro is positioned above both of those releases in Google's model hierarchy and represents the company's direct response to reasoning-focused models from competitors, including OpenAI's o3 and o4-mini series and Anthropic's Claude 3.7 Sonnet with extended thinking.

Key Capabilities

Gemini 2.5 Pro incorporates a built-in reasoning process that allows the model to work through multi-step problems before producing a final response. This approach, sometimes described as chain-of-thought reasoning, is designed to improve accuracy on tasks involving mathematics, coding, and complex logical analysis.

The model accepts text, images, audio, and video as input, maintaining the multimodal design that has characterised the Gemini family since its initial release.

Google reported that Gemini 2.5 Pro achieved top scores on several industry benchmarks at the time of release. On the GPQA Diamond benchmark, which tests graduate-level scientific reasoning, the model scored 84 percent. On the AIME 2025 mathematics competition benchmark, it scored 86.7 percent. On Codeforces, a competitive programming evaluation, it reached a rating of 1260, placing it in the 70th percentile of human competitors.

Google also reported the model ranked first on the Chatbot Arena leaderboard, a crowdsourced evaluation platform maintained by researchers at UC Berkeley where human users compare model outputs directly.

Access and Pricing

Gemini 2.5 Pro is available in preview through Google AI Studio at no cost during the initial period. Paid API access is available through Google Cloud's Vertex AI platform. Google has not announced a general availability date or confirmed final pricing tiers for the model beyond the preview phase.

The model is accessible via the standard Gemini API, allowing developers to integrate it into applications using existing tooling and authentication credentials.

What It Means in Practice

The one-million-token context window allows the model to process roughly 750,000 words, or approximately 30,000 lines of code, in a single request. This capacity is relevant for use cases such as analysing large codebases, processing lengthy legal or scientific documents, or summarising extended transcripts without segmenting the input.

The reasoning capability is designed to reduce errors on tasks that require sequential logic, which has historically been a weakness of large language models responding in a single pass.

Google has not announced a timeline for removing the preview designation or for releasing additional models in the Gemini 2.5 family, though the company has indicated further variants are in development.