Google announced a major upgrade to Gemini 3 Deep Think on July 13, highlighting a strong jump in reasoning benchmarks. The company said the model scored 84.6% on ARC-AGI-2, ahead of Claude Opus 4.6 in Thinking Max mode at 68.8% and GPT-5.2 in Thinking xhigh mode at 52.9%. The source also cited an average human score of about 60%. On the original ARC-AGI-1, Deep Think reached 96%.
Google has opened Deep Think to Google AI Ultra subscribers, while API access is available in early access for enterprise customers. The company’s framing centers on reasoning and scientific use cases rather than general-purpose assistant performance alone.
Benchmark results span reasoning, coding, and exam-style tests
ARC-AGI-2 is described in the source as a benchmark designed to reduce the value of memorized answers and emphasize rule discovery from examples. Beyond the ARC tests, Gemini 3 Deep Think posted 48.4% without tools on Humanity’s Last Exam, a benchmark built by experts across fields to be difficult for AI systems, and Google said that result set a new record.
The update also came with competitive programming and academic testing claims. Google said Deep Think achieved a Codeforces Elo of 3,455, which corresponds to the “Legendary Grandmaster” tier. It also reached gold-medal level on the written portions of the 2025 International Physics Olympiad and International Chemistry Olympiad.
Google points to a math paper review case
Outside benchmark tables, Google cited a research-related example. According to the source, Deep Think reviewed a mathematics paper that had already gone through human peer review and identified a logical flaw that previous reviewers had missed. A mathematician at Rutgers University confirmed that finding. The source did not include the paper title or added technical detail.
That example matters because it shifts the discussion from closed benchmarks to open-ended scientific work. Reviewing a paper is a different kind of task. It depends on consistency, structure, and the ability to catch faults in a setting where there is no simple answer key.
Competition is also about distribution
The source says market share in consumer AI has been moving as the model race intensifies. ChatGPT’s share has dropped from a peak of 87% to about 68%, while Gemini has climbed from under 5% to above 18%. Claude, according to the same source, continues to gain ground in enterprise use.
Google’s position is tied not only to model performance but also to distribution. Gemini is built into Android, Chrome, Google Workspace, and Search. That reach can help adoption quickly, but it also means the product is exposed to a wider group of users who may not have chosen it directly.
AI infrastructure demand spills into crypto narratives
The source links the latest AI model push to rising demand for computing infrastructure. It says the cost of GPU clusters needed to train frontier models has expanded from the hundreds of millions of dollars in 2024 to the billions of dollars by 2026. In crypto, one effect is the shift by bitcoin miners toward AI compute services.
As mining economics tighten, operators with large-scale power and hardware footprints are being pushed to look for contract revenue tied to AI computing. The source, citing a JPMorgan estimate from this week, said BTC production cost had fallen to $77,000 while the bitcoin price was near $66,000. It also noted that AI-related crypto tokens often see short-term speculation after major releases from Google, OpenAI, or Anthropic, even though decentralized compute still faces limits in latency and throughput compared with enterprise AI training needs.