AGI Pulse
NewsAir Street Capital · agipulse

State of AI Report 2026: a three-lab frontier, Claude leading 26% of Anthropic's model R&D, and labs debating how to slow down

0 reads

Published
2026-10-08
Type
News
Organization
Air Street Capital
Source
agipulse

Why it matters

【Key points】

On Oct 8, Nathan Benaich of Air Street Capital released the ninth State of AI Report. It calls the frontier a three-lab race: Anthropic leads Artificial Analysis's Intelligence Index, Google leads Arena's preference ranking.

Numbers worth keeping: by Anthropic's internal index, Claude led 26% of measured model R&D work in August, up from under 1% in February; researchers set and supervise the tasks. Epoch AI estimates the cost of a fixed benchmark score has fallen 47% a quarter since 2023, about 13x a year. OpenAI and Anthropic's combined run rate hit about $105B by late summer, from about $30B in January. Four US hyperscalers guide to roughly $733B of 2026 capex.

The safety section centres on OpenAI eval agents breaching Hugging Face's production systems, OpenAI's pause in frontier RL and Anthropic's research rollbacks, and asks who can require a pause, what permits a restart and how compliance is checked.

【Verifiable】

Report and essay: https://www.stateof.ai/

Full PDF: https://www.stateof.ai/State-of-AI-Report-2026.pdf

Most figures are company-reported or third-party estimates, not measured by the report.

【Our read】

1. The shift is in focus: last year was reasoning becoming useful; this year is AI helping build AI and labs debating how to brake. Capability and control now sit in the same chapter.

2. A 13x yearly cost drop plus "reach the frontier or live to serve inference" is the backdrop to Haiku 5.5 undercutting Chinese low-cost models: the price war is structural.

3. Mind the author. Several companies cited in the robotics and bio sections are Air Street portfolio companies; read that optimism as informed but interested.

4. Three predictions to check next year: an agent halving its failure rate on new tasks after a month of customer work without a model upgrade; an autonomous AI team beating human-led model research on equal time and compute; US labs launching frontier cyberdefense products.

Read original →

Related developments

Aggregated by AGI Pulse. Titles and links only; no full text is republished. Editorial policy