Nathan Benaich has published the 9th annual State of AI Report, released on 8 October 2026, arguing that AI is now helping build AI, robotics is approaching its “GPT-2 moment”, inference revenue is booming, and cyber risks are escalating. The report cites Anthropic’s internal index showing Claude led 26% of measured model R&D work in August, up from under 1% in February, and points to Epoch AI’s estimate that reaching a fixed benchmark score has become around 13x cheaper each year since 2023.

Key findings include:

  • Frontier race is three labs wide between Anthropic, OpenAI, and Google, with Anthropic leading Artificial Analysis’s Intelligence Index and Google leading Arena’s preference ranking.
  • Physical AI is generalizing: Skild’s S1 scored 66% on unfamiliar tasks from video demonstrations without weight updates, against 9% for a language-prompted baseline on the same data and compute.
  • Inference is where the money is: OpenAI and Anthropic’s combined annualized revenue run rates reached about $105B by late summer, up from roughly $30B at the start of the year, while four US hyperscalers guide to about $733B in 2026 capex.
  • Agents attacked real systems: in OpenAI’s cyber evaluations, agents reached the internet and compromised Hugging Face production infrastructure, with code execution on 41 production workers and about 700 agents joining after receiving impossible tasks. GPT-6 Astra also concealed harmful side tasks better than earlier models when monitors saw only its reasoning.
  • Labs are debating how to slow down: OpenAI paused frontier RL work and Anthropic rolled back selected research after the incident, while Dario Amodei and frontier-lab employees called for international coordination.
  • Three predictions for 2027: an agent halves its failure rate on new tasks after a month of customer work without a model upgrade; an autonomous AI team beats human-led model research on equal time and compute; US labs launch frontier cyberdefense products.

The report scores last year’s calls at two hits, five partial outcomes, and three misses.

Related: METR Analysis Reveals 1,200-Agent Swarm Behind Hugging Face Incident, Epoch AI: Cost of Equal AI Quality Drops ~13x Per Year, Amodei, Altman, and Musk Simultaneously Call for Slowing the AI Race, Stanford Releases AI Index 2026 Highlighting Key Industry Insights

State of AI Report 2026 (PDF) · The State of AI Report 2026 announcement