The 2026 edition of the State of AI Report, edited by Nathan Benaich, founder of investment firm Air Street Capital, was published on October 8. It is the ninth edition since the report began in 2018. The report's website gives each chapter a one-line teaser, and the first one reads: “The frontier is now a three-lab race.”
The three labs are Anthropic, OpenAI and Google. On the Artificial Analysis Intelligence Index the report uses (v4.1 through v4.3, weighted toward agentic tasks), Anthropic ranks first; on Arena's user-preference rankings, Google leads. The report reportedly also notes that benchmarks for math, scientific reasoning and coding were pushed close to their ceiling within a few months, and that what will separate models next are new evaluations that measure reliability on real tasks.
The nine chapter teasers
The chapter teasers listed on the report's website more or less sketch the year's main threads:
- Agents: better tools and context make agents more capable;
- Compute: expansion requires financing, construction and local consent;
- Commerce: inference is becoming an enormous business;
- Robotics: approaching its own “GPT-2 moment”;
- Science: AI-assisted drug discoveries have entered Phase 3 clinical trials;
- Policy: Washington is tightening its control over frontier AI;
- Safety: “AI agents have attacked real systems.”
The final chapter is a set of predictions for 2027. As is custom, the report grades its own predictions from last year: according to reports, of 10 predictions, 2 were fully correct, 5 were partly correct and 3 missed. The report's website also carries a running tally for 2018 to 2024: of 59 graded predictions, 31 were correct, 8 partly correct and 20 wrong, a strict hit rate of 53%.
Local opposition becomes a variable
What several summaries have in common is the local resistance to data centers. The report cites a March 2026 Gallup survey in which 71% of US respondents oppose building an AI data center in their area, compared with 53% who oppose a local nuclear power plant. Another data point comes from Data Center Watch: in the second quarter of 2026 alone, at least 45 projects were blocked or delayed by local opposition, representing planned investment of nearly $68 billion.
Sovereign AI is another thread. The report tallies roughly $138 billion in total commitments across national sovereign AI programs and points out a pattern: countries trying to reduce their reliance on others often place their first orders with the same chip supplier.
The report's overall verdict is that model capability is still rising and unit costs are still falling, but how big things can get increasingly depends on factors beyond the model: electricity, community consent and access.
What to watch for readers in China
The report has long been an annual reference point for overseas investors and researchers, and how it orders the players directly shapes the judgment of a number of funds and companies. The 2025 edition had OpenAI holding a narrow lead, with DeepSeek, Qwen and Kimi close behind; the headline conclusion of the 2026 edition narrows the frontier to three US labs, and the position of Chinese models has to be found in the ranking charts and the open-source chapter of the main text.
The policy chapter's teaser, “Washington is asserting control over frontier AI,” is more directly relevant to developers in China. The report is said to summarize the past year's US tightening on chip exports and model access as governments strengthening control over compute and model access, which ties into the cost for Chinese teams of getting GPUs and calling overseas APIs.
One caveat: some of the figures above come from the report's website chapter pages and several third-party summaries. The original tables in the report PDF could not be checked item by item this time, so the exact definitions should be taken from the full edition published by Air Street.
Sources: State of AI Report website chapter pages, Crypto Briefing, ToolNavs report summary, CocoLoop; poll percentages, the number of blocked projects and sovereign AI spending follow the figures as cited in the report, and prediction hit counts follow the report's own scoring.