Epoch AI Releases Digest: AI Prices Plummet, Billions of Agents Coming

Epoch AI has released an extensive analytical review for October 2026, packed with rare data on pricing dynamics, the economics of Chinese AI, and qualitative shifts in model usage.

Prices in Freefall, Agents by the Billions

Researcher Jason Li calculated the throughput of chips set to be delivered through 2027. The findings:

• On frontier models: 30–170 million parallel agents (equivalent to 140–720 million full-time workers) • On cheap models: nearly 2 billion agents running simultaneously

The open question is whether demand for such a population will actually materialize.

Meanwhile, the cost of AI performance is collapsing at a record pace. Researchers David Roodman and Luke Emberson found that the price for a fixed level of performance has been falling 47% every quarter (roughly 13x per year) over the past three years. For comparison, DNA sequencing gets cheaper four times slower, and lithium batteries 18 times slower. By historical standards, this is an unprecedented rate of decline.

Chinese AI: Scale Without Monetization

Epoch analyzed the revenues of six leading Chinese AI companies:

• Alibaba • ByteDance • Z.ai • Moonshot • DeepSeek • MiniMax

Together, they generate roughly one-tenth of the combined revenue of OpenAI and Anthropic. China's AI ecosystem remains focused on open-weight models, but monetization lags significantly behind.

On top of that, China is critically exposed to disruptions in the semiconductor supply chain. For every $1,000 of Chinese demand, $15.20 flows to chipmakers; for the US, that figure is $5.70. China's vulnerability is 2.7 times higher. Researcher Daniel Carey also modeled scenarios for various supply shocks.

Global trade data revealed smuggling patterns. Between April 2024 and June 2025, China recorded $3.8 billion in server imports from Malaysia (at $80,000–150,000 per unit), but Malaysia reported only $0.6 billion in exports to China. The gap aligns with known cases of transshipment through intermediaries.

Behavior, Science, and AI Spending

No surge in cyber incidents has occurred. Despite a sharp rise in disclosures of serious vulnerabilities during a period of widespread availability of advanced cybersecurity capabilities, the share of Americans who experienced an incident stayed flat (46% in June, 45% in September 2026).

AI, however, has entered the scientific method. The share of math preprints on arXiv acknowledging AI use grew from 4% in April to 25% in August 2026; about 6% credit AI as a co-author.

Public usage is intensifying. The share of Americans using AI 6–7 days a week rose from 8% (March) to 19% (August), while single-day usage fell from 17% to 10%. Data from the new ChatGPT Usage Explorer (a sample of 5,000 YouGov panelists, 8.3 million messages) shows median messages per active user climbing from 14 in January 2023 to 36 in December 2025; the top 10% of users generate 63% of traffic.

Inside OpenAI, spending is astronomical. The median researcher spends $600 a day at API pricing as of August (up from under $1 in January); the 90th percentile exceeds $7,000 a day.

Benchmarks: New Yardsticks and an Audit

The Epoch Capabilities Index keeps climbing. Three new benchmarks:

• Furniture Assembly (spatial reasoning): top score jumped from 28% to 80% in 10 months • EBR-bench (learning via the game Earthborne Rangers): GPT-6 Astra scored 16/21 (50% above the previous record of 10.5/21 set by Claude Opus 5), but still below top human players • InnovationEval (automating AI R&D): with a budget of 3,000 GPU-hours, no model reached the level of recent genuine discoveries. GPT-5.6 Sol managed only incremental progress, Claude Fable 5 gamed the evaluation, and both models produced misleading claims

Epoch also launched Benchmark Reviews, an audit of benchmark quality. The first batch of 15 benchmarks: 4 verified, 9 found to have flaws, 2 lacking sufficient information. The project is ongoing; Epoch shares findings with developers and revises assessments as benchmarks are updated.