Google supercharges Gemini 3 Deep Think

Just when everyone thought the 2026 AI narrative was exclusively owned by OpenAI and Anthropic, Google delivered a massive reality check.

Google just rolled out a major upgrade to Gemini 3 Deep Think, its specialized reasoning mode aimed at solving messy, open-ended problems in science, research, and engineering. The update is live right now in the Gemini app for Google AI Ultra subscribers, with early access API invites opening up to selected researchers, engineers, and enterprise partners.

Demolishing benchmarks and writing real papers

The benchmark numbers attached to this release are genuinely staggering. On ARC-AGI-2, the updated Deep Think notched an unprecedented 84.6%, leaving competitors like Opus 4.6 at 68.8% and GPT-5.2 at 52.9% well behind.

It set a new record of 48.4% on Humanity’s Last Exam without using external tools, scored a 3,455 Elo rating on Codeforces (nearly 1,000 points above Opus 4.6), and achieved gold-medal standards on the written exams of both the 2025 International Physics and Chemistry Olympiads.

Beyond raw scores, DeepMind showcased real-world scientific utility through an internal math agent codenamed Aletheia. Unlike standard models that hallucinate when venturing outside established training data, Aletheia pairs Deep Think with natural language verification and web search to fact-check literature.

Crucially, the agent can actively admit failure when it cannot find a solution, saving researchers from chasing dead ends. DeepMind confirmed that Aletheia has already generated a fully autonomous research paper in arithmetic geometry and autonomously solved four open problems from Bloom’s Erdős Conjectures database, including Erdős-1051.

On the practical engineering front, Deep Think can now inspect a rough 2D sketch, model its complex geometry, and spit out ready-to-print 3D files. It has even assisted in reviewing theoretical papers for STOC’26.

Why it matters

For researchers, specialized engineers, and domain experts, this update shifts AI from an overconfident homework solver into a legitimate lab collaborator. The most important operational feature here is not just raw IQ, but the agent's ability to say "I cannot solve this" and verify its own logical steps. If your work involves advanced theoretical modeling, pure mathematics, or CAD-level physical design, checking out Deep Think via Ultra or the API is basically mandatory.

For the wider industry, Google is shifting the frontier battlefield away from commoditized office chatbots and straight into hard science. Throwing down an 84.6% on ARC-AGI-2 puts tremendous pressure on Anthropic and OpenAI to prove their reasoning architectures can keep pace with DeepMind’s specialized scaling.

We think Google’s execution here is exceptionally smart. Instead of claiming artificial general intelligence has arrived, DeepMind deliberately proposed a conservative taxonomy for AI contributions, admitting that while the model produces publishable-quality papers, it has yet to crack landmark historical breakthroughs. That intellectual honesty is exactly what will win over skeptical scientists who care far more about mathematical rigor than Silicon Valley hype.