Google has quietly flipped the switch on “Deep Think,” its latest and most ambitious reasoning engine, giving Google AI Ultra subscribers first dibs on what the company hopes will redefine how we tackle complex problems. First teased at Google I/O back in May, Deep Think is now rolling out in the Gemini app to anyone paying the premium $249.99-per-month AI Ultra subscription—no extra download required.

Deep Think isn’t just another incremental upgrade. Under the hood, it embraces a “parallel thinking” approach that spawns multiple AI agents to explore different angles of a question at once—then pits those ideas against each other to pick the strongest response. In effect, it mimics a brainstorming session in software form, juggling a handful of reasoning paths simultaneously rather than marching down a single thought process. Google claims this lets Deep Think handle choreography-level complexity: think multi-step math proofs, intricate coding bugs, or strategic planning tasks that would flummox a one-track AI model.

Perhaps most eye-catching is Deep Think’s recent showing on the world stage: a modified variant of this very engine snagged a gold medal at the 2025 International Mathematical Olympiad when tested by a small cohort of academic collaborators. The public version you’ll play with today has been dialed back for speed—enough to hit bronze-medal standard in internal trials—but is said to be razor-sharp compared to anything in the consumer market. While Google’s blog teases that the full gold-medal model “takes hours to reason,” the Ultra-tier release wraps that power into a form that delivers answers in seconds.

Benchmark tests back up the hype. In head-to-head trials on LiveCodeBench V6 and the cheekily named Humanity’s Last Exam, Deep Think outscored both OpenAI’s o3 model and xAI’s Grok 4—previously considered top dogs in heavy-lifting reasoning tasks. And according to Google, its evaluation suite also covered real-world coding challenges and scientific question sets, with Deep Think emerging as the clear leader. That combination of raw reasoning prowess and real-world chops is exactly what Google is betting will give Gemini a leg up on rivals.

Of course, such firepower doesn’t come cheap—or wide. Only AI Ultra subscribers unlock Deep Think in the Gemini app (or on the web), and Google has capped usage at a fixed—but undisclosed—number of prompts per day. The $249.99 monthly price tag includes expanded model context windows, priority access to future AI experiments, and extra cloud storage, but it still puts Deep Think out of reach for casual tinkerers. Google’s lower-tier AI Pro plan—formerly AI Premium at $20 per year—remains a more wallet-friendly option, though it lacks the cutting-edge reasoning engine.

Looking ahead, Google may well broaden Deep Think access beyond the Ultra circle. Early trusted testers have already helped iron out performance kinks, and precedent suggests that once the model proves stable in the field, it could trickle down to lower tiers—or even become available via API for developers. But for now, Deep Think remains a premium perk, an engine room reserved for the most daring AI explorers.

For those already on board the AI Ultra train, Deep Think lands today in the Gemini app—ready to tackle your toughest puzzles. And for the rest of us, it provides a tantalizing glimpse at how far language models can go when they stop thinking alone and start thinking together.