Gemini 2.5 Deep Think lands in the Gemini app for AI Ultra subscribers
Google's parallel-thinking reasoning mode goes live behind a paywall, with a Vertex AI track for trusted testers and a 32K thinking-budget knob for enterprises.
Google began rolling out Gemini 2.5 Deep Think on June 22 to Google AI Ultra subscribers in the Gemini app, with a parallel track for trusted testers on Vertex AI. The pricing tier is the headline: parallel reasoning is now an explicitly paywalled capability, sitting one rung above the standard 2.5 Pro experience.
In the app, the interaction model is deliberately mundane. Users pick 2.5 Pro from the model drop-down, toggle “Deep Think” in the prompt bar, and wait. Responses generally take “a few minutes,” after which the app pings the user. This is a chat product that has stopped pretending to be conversational.
Deep Think generates multiple hypotheses in parallel, revises and combines them, and trades inference time for solution quality. Google says novel reinforcement-learning techniques “encourage the model to make use of these extended reasoning paths.” The shipped build is the same one Google previously ran past “a small group of mathematicians and academics” at what it calls the gold-medal standard. Google claims state-of-the-art results, without tool use, on LiveCodeBench V6 and Humanity’s Last Exam.
The safety profile is more candid than usual. Google notes Deep Think shows “improved content safety and tone-objectivity” versus 2.5 Pro, but also “a higher tendency to refuse benign requests.” That’s the well-documented tax on heavier deliberation.
For enterprises, the Vertex AI path matters more than the app demo. Thinking Budgets are configurable up to 32K tokens, giving buyers a literal dial between latency and depth. Google Cloud is also pitching 2.5 as its “most secure model family to date” against indirect prompt injection during tool use, language aimed squarely at the agent-deployment anxieties that have shadowed every enterprise pilot since the 2024 prompt-injection disclosures.
Deep Think arrives a month after Gemini 3.5 Flash, which Google launched at its May 19 developer conference as its most agentic model yet. The shape of the lineup is now legible: Flash for agents, Deep Think for the hard problems someone is willing to pay per-minute to solve. Chat is no longer the product.
Sources
- https://blog.google/products/gemini/gemini-2-5-deep-think/
- https://ai.google.dev/gemini-api/docs/changelog
- https://gemini.google/release-notes/
- https://cloud.google.com/blog/products/ai-machine-learning/expanding-gemini-2-5-flash-and-pro-capabilities
- https://techcrunch.com/2026/05/19/with-gemini-3-5-flash-google-bets-its-next-ai-wave-on-agents-not-chatbots/