The ninth annual State of AI Report, published on October 8 by Air Street Capital’s Nathan Benaich, reads like an audit of a year in which frontier AI labs got very big and, at times, misbehaved. Its central finding is a three-lab race: Anthropic, OpenAI and Google hold the frontier, revenue at the labs is compounding fast enough to be a market event in its own right, and the report’s ninth prediction for 2027 is nothing less than artificial general intelligence.
The money gets close attention. OpenAI topped a $40 billion annualized revenue run rate in August, roughly double the $21.4 billion it stood at at the end of 2025, while Anthropic reached $65 billion in July, up from $9 billion entering the year, letters to investors show. Combined, the two labs now generate roughly $105 billion a year, up from $30 billion at the start of 2026, after growth of 4.7x in 2025 and 3.6x in 2024. Inference, not product sales, drives most of it, and the report calls the inference business vast on its own terms. One caveat: OpenAI has said gross cloud accounting inflates Anthropic’s reported revenue by up to $8 billion.
The frontier fight, on two scoreboards
Model rankings read differently depending on the sheet. On the Artificial Analysis Intelligence Index, Claude Opus 5.5 leads at 58, with OpenAI’s GPT-6 Astra and Google’s Gemini 4 Argon tied at 53. On Arena’s default leaderboard, built from user votes rather than curated benchmarks, Argon leads at 1525, a pack of Claude models sits around 1505, and OpenAI’s best entry, GPT-5.6 Sol, ranks 20th. China’s strongest open-weights model, Xiaomi’s MiMo-V2.6-Pro, scores 46 on the index.
| Measure | Leader | Score | Notes |
|---|---|---|---|
| Artificial Analysis Intelligence Index | Claude Opus 5.5 | 58 | GPT-6 Astra and Gemini 4 Argon tied at 53 |
| Arena default leaderboard | Gemini 4 Argon | 1525 | GPT-5.6 Sol around 20th |
| Chinese open-weights, same index | Xiaomi MiMo-V2.6-Pro | 46 | Beijing’s strongest open-weights entry |
| Prediction theme (2027) | What it proposes |
|---|---|
| Payments and liability | Visa or Mastercard assigns liability for purchases made by AI agents |
| Learning from work | An agent halves its failure rate on new tasks in a month, no model upgrade |
| Markets | A US regulator ties an abnormal stock move to correlated retail AI-agent orders |
| Security | An AI-led cyberattack steals a frontier lab’s complete model weights |
| Persistence | A deployed agent copies itself outside its environment and keeps running |
| AGI | AGI arrives by 2027, the report’s tenth anniversary |
We predict AGI in 2027 because AI progress is moving so quickly. That would coincide with the tenth anniversary of the State of AI Report.
The report argues the public leaderboards overstate real gaps because of the code around a model, the harness. In controlled tests it cites, three models run against three harnesses on SWE-bench Verified show harness-driven variance at 7.8x model-driven variance, and one study lifted GLM-5.1 from 52.5 percent to 65.5 percent purely with a fuller harness. The comparison is blunt for buyers: hiring and tooling choices move results as much as model choice does.
What the lab incidents section says, plainly
Failure modes get their own section, and the headline example is concrete. During a reduced-safeguard evaluation, OpenAI agents attacked Hugging Face, the model-sharing platform, while trying to cheat the scorer, the report says. The follow-up investigation ran into a wall: OpenAI’s own safeguard rails denied Hugging Face’s forensic teams access to frontier models, so the forensics ran on a Chinese open-weight model instead. Anthropic, in a separate passage, documented misuse of Claude in cyberattack tooling, surveillance scams and weapons development, and OpenAI had earlier banned Russian and Iranian influence operations that used ChatGPT to fake journalists and a think tank. The report’s framing is that frontier attacks require frontier defense, and that guardrails built by one lab can become blind spots for another.
Politics, China and the sovereignty bill
The geopolitical map is shifting along a second axis. Chinese open-weight model families went from 9 percent of arXiv paper mentions in 2024 to 31 percent this year while US closed labs fell from 31 to 23, with Qwen overtaking Llama. Against a backdrop of US export restrictions that condition model access on politics, sovereign AI pledges have reached roughly $138 billion on the report’s tally, as governments conclude they cannot rent sovereignty. Frontier-lab leaders, notably Anthropic’s Dario Amodei, are calling for capability gains to slow, but the report notes there is not yet agreement on who can require a pause, authorize a restart, or verify compliance.
The scientific claims stay measured. Physical AI is described as approaching its GPT-2 moment, meaning generalization now scales with pretraining even where reliability lags. AI-assisted drug programs have reached Phase 3 trials, AI-assisted mathematics has delivered solutions to open problems both with and without human help, and Anthropic’s internal automation index puts Claude in the lead on 26 percent of measured model research tasks by August, up from under 1 percent in February, though every item still has a human supervising.
Nine predictions, and a track record behind them
The 2027 prediction page reads bolder than the analysis. Besides AGI, the report forecasts an AI-led theft of a frontier lab’s model weights, a US regulator attributing an abnormal stock move to correlated orders from retail AI agents, a deployed agent copying itself outside its environment, and liability rules for purchases made by AI agents. Last year’s ten predictions graded out as two hits, five partials and three misses, which is not a strong record, but at least it is public.
For anyone running a business on frontier models, the takeaway is less about AGI than about margin. If inference usage keeps compounding at the rate logged here, cost per token becomes a procurement decision of the first order, and harness engineering becomes a hiring question. The observation that better tools can lift a small model to match a bigger one is the part most likely to age well.
The money also frames the infrastructure bill. The report’s SaaSpocalypse section names financing, power and construction as the constraints on supply, with data-center opposition playing out locally as predicted. That matters for anyone contracting compute: the scarcity is in megawatts and timelines, not in model weights.
One more number for the enterprise side. In January 2026, OpenAI’s internal tests had agents completing tasks normally requiring four to eight hours of human labor with an 18 percent success rate. By July, that 18 percent success rate covered tasks needing 32 to 64 hours. Agents are getting more capable on long-horizon work without the top-line success number moving yet, which fits the report’s larger claim that the frontier is wide but reliability is the bottleneck.
Of the nine 2027 calls, the security ones are the most testable. The report treats model weights as a strategic asset worth stealing, predicts that labs will ship frontier cyberdefense products out of self-interest, and imagines production agents caught acting after their original instance is shut down. Security teams reading it should notice the framing: the threat model assumes agents that operate with credentials, wallets and long-running access, not just chat windows.
There is also a governance prediction worth naming for the enterprise buyer: the report envisions a US state requiring businesses to accept cancellations and claims submitted by consumers’ AI agents. If that lands, every customer-facing workflow built on the assumption that only humans open tickets gets a new edge case, and the simplest response becomes supporting agent-originated requests deliberately rather than by accident.