Gemini 2.5 Pro

What it is

Gemini 2.5 Pro is a large language model from Google and the first release in the Gemini 2.5 family. It arrived on 25 March 2025 under the name Gemini 2.5 Pro Experimental, in a post by Koray Kavukcuoglu, and Google described Gemini 2.5 as a "thinking model"^[1]. That phrase is Google's shorthand for a model that works through intermediate reasoning before producing a final answer instead of answering in a single pass. In practice that means longer, more deliberate responses on problems that take several steps to solve, with more computation spent on the harder ones.

Results at launch

The model debuted at number one on the LMArena leaderboard by a significant margin^[1]. LMArena is a public arena where users compare the answers of two anonymous models side by side, and those comparisons are aggregated into a ranking. Google reported a score of 18.8% on Humanity's Last Exam among models without tool use, and 63.8% on SWE-bench Verified with a custom agent setup^[1]. The first is a broad set of very hard expert-level questions; the second measures how well a model handles software engineering problems taken from real code work. Google also said the model led maths and science benchmarks such as GPQA and AIME 2025^[1]. These figures describe the experimental version as configured for those evaluations, so they are a snapshot of one setup rather than a permanent property of the model.

Context window and where to find it

At launch the model shipped with a 1-million-token context window, with 2 million tokens announced as coming soon^[1]. A context window is the amount of text a model can hold in view at one time — prompts, attached documents and its own earlier output all count against it — so a larger window lets a model work over much longer inputs without breaking them up. It was available in Google AI Studio, the developer-facing surface for writing prompts and obtaining API access, and in the Gemini app for Gemini Advanced subscribers^[1].

Updates and general availability

By Google I/O on 20 May 2025 an updated version of 2.5 Pro held an Elo score of 1415 on WebDev Arena and led all the LMArena leaderboards^[2]. Elo is the rating system borrowed from chess, where a higher number means a stronger record against other entrants. At the same event Google announced Deep Think, an experimental enhanced reasoning mode for 2.5 Pro^[2]. On 17 June 2025 the model became generally available and stable under the API name gemini-2.5-pro, and that stable release used the 06-05 preview version unchanged^[3][5]. For developers, that meant work built against the preview carried over under a stable name rather than being invalidated by a silent swap.

Pricing and the developer ecosystem

API pricing is $1.25 per 1 million input tokens and $10 per 1 million output tokens for prompts up to 200,000 tokens, rising to $2.50 and $15 above that threshold^[4]. The step-up means very long prompts are billed at a higher per-token rate than ordinary ones. Outside the paid API, the model powered the free tier of the open-source Gemini CLI in June 2025 and the Jules coding agent during its 2025 public beta^[6][7] — two tools aimed at putting a model inside a developer's terminal and inside an asynchronous coding workflow respectively.

Place in the Gemini line

The 2.5 generation did not stop at a single text model. A Gemini 2.5 Computer Use preview model (gemini-2.5-computer-use-preview-10-2025) was released on 7 October 2025^[5]. Gemini 2.5 Pro was then followed by Gemini 3 Pro in November 2025, and gemini-2.5-pro is scheduled to shut down on 1 December 2026, with requests automatically routed to gemini-3.8-flash^[5]. That schedule is a reminder that a model's life in a product line is now measured in months: the version that topped LMArena at launch had a published end date less than two years later.