The State of AI Report 2026 was published on October 8, 2026 by Nathan Benaich, founder of Air Street Capital, via Air Street Press. Now in its ninth edition, the report sums up the year in three themes: the frontier race has narrowed to three labs, agents are producing value in real work, and governments are tightening control over compute and model access.
There Is No Longer a Single Name in First Place
The report argues that frontier competition is now a three-way contest between Anthropic, OpenAI, and Google, and that different leaderboards crown different winners: Anthropic leads on Artificial Analysis's Intelligence Index, while Google leads Arena's ranking of the answers people prefer. Meanwhile, the benchmarks once used to separate the labs are saturating fast; in math, scientific reasoning, and coding, some tests approached their ceilings within months. What will distinguish models next, the report suggests, is whether new evaluations can measure how reliable a model is on real tasks.
Agents Are Changing How Research Itself Gets Done
One example comes from Anthropic's internal data: by August 2026, Claude led 26% of the company's measured model R&D work, up from under 1% in February of the same year. At the same time, OpenAI adoption research finds the fastest agent growth not among developers but among non-technical roles in legal, sales, recruiting, and marketing. The report's judgment is that the gap between people who get a lot from AI and those who get little is now mostly a matter of how the tools are set up and used, not of model capability alone.
Costs Are Falling, Bills Are Rising
Citing Epoch AI estimates, the report notes that since 2023 the inference cost of reaching a fixed benchmark score has fallen about 47% per quarter, roughly 13 times cheaper each year. Demand is growing faster still: OpenAI and Anthropic's combined annualized revenue run rate reached about $105 billion by late summer 2026, up from about $30 billion at the start of the year. Supply-side spending is just as striking, with four US hyperscalers guiding to roughly $733 billion in combined 2026 capital expenditure.
Data Centers Are Starting to Lose at the Community Gate
The report's cooling-off section looks beyond supply: a Gallup survey from March 2026 found 71% of Americans opposed a local AI data center, a higher share than the 53% who opposed a nuclear plant. Data Center Watch counted at least 45 projects, representing nearly $68 billion in planned investment, blocked or delayed by local opposition in the second quarter of 2026 alone. Sovereignty is the other thread: sovereign AI programs in the report pledge about $138 billion in total, while NVIDIA alone reported more than $30 billion in sovereign AI revenue in fiscal 2026. Countries trying to escape dependence often end up ordering from the same supplier first.
The report matters because it puts three facts on one table: model capability is still rising and unit costs are still falling, but the ceiling is increasingly set by things outside the model, namely power, community consent, and access rights. The practical takeaway for builders is concrete: do not pick models from a single leaderboard, and do not budget from token prices alone without accounting for the call volume that agents bring with them.
