Deep Think Arrives: Google Quietly Drops Gemini 2.5 Pro to Seize the AI Crown

The AI arms race has a new king. Google has surprise-released its most capable model to date: Gemini 2.5 Pro with Deep Think reasoning mode. In a single move, the tech giant has completely rewritten the industry leaderboards.

Designed to tackle highly complex reasoning tasks, this model is not just a minor iteration. It is Google’s definitive answer to high-compute, extended-inference AI models, offering unprecedented power for developers and enterprise teams alike.

Double the Context, Triple the Reasoning

The absolute headline of this release is twofold: the massive 2 million token context window and the game-changing Deep Think reasoning mode.

• 2 Million Token Context: Google has doubled the working memory of the previous Pro model. Gemini 2.5 Pro can now seamlessly ingest and reason over massive datasets, entire software codebases, hours of high-definition video, or several full-length textbooks in a single prompt session.
• Deep Think Mode: Inspired by advanced reinforcement learning and parallel hypothesis testing, Deep Think allows the model to "pause" and spend significantly more compute on complex problems before returning a response. This stops the AI from drifting into confident hallucinations, allowing it to systematically work through intricate logic, math, and coding bottlenecks.

Benchmark-Shattering Results

Google’s internal evaluations and early developer benchmarks paint a staggering picture of Gemini 2.5 Pro’s superiority:

• GPQA Diamond (Graduate-Level Science): A stunning 82.4%, comfortably eclipsing rival models' scores of 79.1% and 76.3%.
• MMLU-Pro: 89.8%, the highest score of any publicly available AI model.
• HumanEval+ (Coding): A record-breaking 94.1%.
• MATH-500: A near-perfect 97.2%.

While other models still hold marginal leads in pure software-agent workflows, Gemini 2.5 Pro stands alone at the peak of science, mathematics, and long-form multimodal reasoning.

Pricing & Availability

For developers eager to build, Google is making the model immediately available. You can access Gemini 2.5 Pro with Deep Think today via the Gemini API, Google AI Studio, and Vertex AI.

Standard mode is priced incredibly competitively at $2.50 per million input tokens and $15 per million output tokens. Utilizing the Deep Think reasoning mode scales the compute footprint, running at roughly 4x the standard rate—a premium price well worth it for solving humanity's most complex technical challenges.

With billions of Android devices and Chrome browsers globally, Google has the distribution pipeline to make Deep Think the default workflow of the future. The benchmark crown has officially been seized.