Gemini 3 Deep Think、科学・研究・エンジニアリングを加速 Gemini 3 Deep Think: Advancing science, research and engineering

Google DeepMind has introduced Deep Think, an enhanced reasoning mode built on top of Gemini 3 Pro and aimed squarely at frontier problems in mathematics, the sciences and software engineering. The release positions Gemini not just as a conversational assistant but as a serious tool for researchers tackling problems that demand sustained, structured thought.

At the technical core, Deep Think extends inference-time computation through what DeepMind describes as parallel thinking. Rather than committing to a single chain of reasoning, the model explores multiple hypotheses and solution paths simultaneously before converging on an answer. The company reports that this approach yields substantial gains over the standard Gemini 3 Pro on demanding evaluations such as Humanity's Last Exam and ARC-AGI-2, as well as on competition-level mathematics and coding tasks. The implication is that for problems where careful deliberation matters more than fast turnaround, Deep Think can reach answers previously out of reach for general-purpose models.

The target audience is explicit: scientists, engineers and advanced developers who need a model capable of grinding through proofs, derivations, complex codebases or multi-step research questions. Access is being rolled out through the Gemini app for Google AI Ultra subscribers, signalling that DeepMind sees Deep Think as a premium capability rather than a default behaviour.

It is worth situating this launch within DeepMind's broader portfolio. The lab has a strong track record of specialised scientific systems, from AlphaFold's protein structure prediction to AlphaProof and AlphaGeometry's medal-level performance at the International Mathematical Olympiad. Deep Think appears to be an attempt to bring that flavour of deep, deliberate reasoning into a general-purpose foundation model, narrowing the gap between bespoke scientific AI and an everyday assistant.

The wider industry has been converging on similar ideas. OpenAI's o1 and o3 reasoning models, Anthropic's extended thinking modes for Claude, and xAI's reasoning variants of Grok all rely on spending more compute at inference time to crack problems that defeat conventional decoding. Deep Think is Google's most explicit answer in that race, and the benchmark numbers suggest the gap among frontier labs on hard reasoning tasks remains tight.

There are caveats worth noting. Deep Think is likely to be considerably slower and more expensive per query than the base Gemini 3 Pro, given the parallel exploration and extended reasoning involved. Users will probably need to be selective about when to invoke it, reserving the mode for problems where accuracy genuinely outweighs latency and cost. As with any reasoning-heavy system, careful evaluation in domain-specific workflows will be needed before treating its outputs as authoritative, particularly in scientific contexts where subtle errors can propagate.

Still, the direction of travel is clear. With Deep Think, Google is making a concrete bet that general-purpose LLMs, given enough room to think, can become genuine collaborators in research and engineering rather than merely productivity aids.