Google drops Gemini 2.5 Pro with built-in thinking

While the rest of the AI industry has been busy treating deep reasoning like a specialty feature you pay extra to toggle on, Google is quietly turning it into default equipment.

Google just introduced Gemini 2.5, introducing it as the company's most intelligent AI family yet. The initial release is Gemini 2.5 Pro Experimental, and it didn't just crawl into the arena—it debuted straight at number one on the LMArena leaderboard by a notable margin.

Built-in thinking and heavy-duty benchmarks

Previously, Google dabbled in reasoning models with Gemini 2.0 Flash Thinking. With Gemini 2.5, Google overhauled the underlying base model and refined its post-training to build chain-of-thought analysis directly into the system. The plan is to bake this reasoning behavior into every model across the board moving forward.

On raw evaluations, the improvements are clear. Operating without compute-multiplying tricks like majority voting at test time, Gemini 2.5 Pro took the lead on math and science tests including GPQA and AIME 2025. It also posted an 18.8% score without tools on Humanity's Last Exam, a benchmark assembled by hundreds of subject matter experts to test the absolute outer limits of human knowledge.

Coding saw an equally noticeable jump. Gemini 2.5 Pro scored 63.8% on the SWE-Bench Verified benchmark using a custom agent setup, and 68.6% on Aider Polyglot. According to Google, it is particularly tuned for producing functional web apps and agentic code setups, demonstrating the ability to turn a one-line prompt into fully executable video game code.

As with previous Gemini models, long context remains a focal point. Gemini 2.5 Pro launches with a 1M token window, with Google planning to expand that to 2M tokens in the near future. The model processes text, audio, images, video, and complete source repositories in one place. It is currently available in Google AI Studio and for Gemini Advanced subscribers on web and mobile, with Vertex AI availability and API pricing slated for the coming weeks.

Why it matters

Google just did something very clever: instead of positioning reasoning as a gated, slow, elite model class, it is treating "thinking" as the standard baseline for everyday computing.

If you build software or wrangle giant codebases, this release is worth testing immediately in Google AI Studio. Combining a 1M token window with leading agentic coding scores directly tackles the headache of repo-level modifications. You no longer have to pick between a model that can read your whole application and a model that is smart enough to actually fix it.

That said, questions remain around latency and economics. Reasoning tokens take time, and in casual workflow settings, having a model pause to deliberate can feel sluggish. Furthermore, because this is an experimental release, Google has not yet shared final API pricing, meaning we don't yet know how cost-effective running these thinking loops will be at production scale.

Still, Google clearly grabbed the AI crown back for now. It puts immediate pressure on OpenAI and Anthropic to show that their next-generation reasoning architectures can match both this depth of logic and massive multi-modal context in a single package.