Google brings Deep Think to the Gemini app for complex reasoning

When faced with an impossible problem, humans usually pause, draft three different hypotheses on scratch paper, and pit them against each other until one wins.

Now Google is giving Gemini the exact same luxury. On Friday, the company rolled out Gemini 2.5 Deep Think inside the Gemini app, exclusively available to subscribers of its $250-per-month Google AI Ultra tier.

First previewed at Google I/O in May, this is Google’s first publicly available multi-agent model. While a gold-medal-winning variation tested at this year's International Mathematical Olympiad (IMO) took hours to ponder a single problem, this release is tuned for daily use. According to Google, it still hits Bronze-level performance on the 2025 IMO benchmark, while the original multi-hour gold-medal model is being reserved for select mathematicians and researchers.

Parallel brainstorming with tool use

Unlike traditional reasoning models that chew through a single chain of logic sequentially, Deep Think relies on parallel thinking techniques. The system generates multiple ideas at once, analyzes them simultaneously, and refines or merges different hypotheses before spitting out a verdict.

DeepMind developed novel reinforcement learning techniques to encourage the model to actively explore these extended reasoning paths. It also hooks up automatically with Google Search and code execution, churning out significantly longer and more detailed responses.

On raw benchmarks, the performance jump is hard to ignore. On the notoriously tough Humanity’s Last Exam without tools, Gemini 2.5 Deep Think scored 34.8%, beating xAI’s Grok 4 at 25.4% and OpenAI’s o3 at 20.3%. On competitive coding benchmark LiveCodeBench V6, it posted 87.6%, outpacing Grok 4’s 79% and o3’s 72%.

There are clear trade-offs, though. Google noted that while Deep Think improved in tone objectivity and content safety, it also developed a higher tendency to refuse benign requests. On top of that, Ultra users only get a fixed set of prompts each day.

Why it matters

We think Deep Think confirms that frontier AI is officially in its multi-agent reasoning era. Between xAI's Grok 4 Heavy, Anthropic's research agents, and OpenAI's unreleased IMO architectures, every major lab is betting that throwing massive inference compute at parallel logic paths is the best way to break reasoning bottlenecks.

If you are a competitive programmer, an algorithm designer, or an academic wrestling with tough mathematical conjectures, having an AI that genuinely checks its own work from multiple angles is huge. But for virtually everyone else, the barrier to entry is brutally high.

At $250 a month with daily usage caps, the economics of multi-agent inference clearly remain brutal. Google's rollout strategy proves that while parallel thinking works, serving it at scale is still a massive financial strain. Until these reasoning clusters become cheap enough for ordinary developer APIs, Deep Think will remain an elite playground rather than an everyday workhorse.