State of AI 2026: frontier narrows to three labs, inference costs drop 13x a year

Nathan Benaich (Air Street) · rss · 2026-10-08

Air Street Capital's Nathan Benaich released the 9th annual State of AI Report. Key takeaways:

Frontier is a three-lab race between Anthropic (top of Artificial Analysis Intelligence Index) and Google (top of Arena preference ranking), with OpenAI alongside.
Spiky capabilities, tool dividends: benchmarks near ceilings within months; adding tools/context/feedback lifted GLM-5.1 from 52.5% to 65.5% on 100 SWE-bench Verified tasks without weight changes.
Agent adoption is fastest among non-developers (legal, sales, recruiting, marketing); the value gap is mostly a "human skill issue," argues Benaich.
AI building AI: Anthropic reports Claude led 26% of measured model R&D work in August, up from <1% in February; Karpathy's autoresearch automates overnight experiment loops, though full recursive self-improvement remains undemonstrated.
Robotics' GPT-2 moment: Skild S1 scored 66% on unseen tasks after 100k hours of video pre-training (vs 9% baseline); Wayve GAIA-4 and Odyssey-3 world models cut needed demonstrations.
AI for science: 9 Epoch open math problems solved (4 autonomous); AI drug discovery reaches Phase 3 (Generate's GB-0895, Insilico's rentosertib), with new data-licensing business models (Tempus, GSK-Noetik).
Inference waterfall: Epoch estimates fixed-score cost falls 47% per quarter since 2023 (13x/year), fueling recurring agent-driven inference demand.

More from AGI Musings