Epoch AI is a research organization that studies trends in artificial intelligence, including model capabilities, costs and usage. Recent coverage centers on its findings: GLM 5.3 building working Firefox exploit chains in under 48 hours at roughly $1,000 in API fees each; a joint look with Anthropic finding agents reach at best 15% of the human reference score; estimates that inference prices at fixed performance fall about 13x per year; and an Ipsos survey showing near-daily AI use among US adults rising from 8% to 19%.
Open-weight GLM 5.3 builds Firefox exploit chains autonomously in under 48 hoursEpoch AI researchers report that the publicly available GLM 5.3 converted public security fixes into working exploits, compromising six Firefox releases before patches reached users at roughly $1,000 in API fees each, without tripping any provider's safeguards.
China's Open-Weight Models Trail US Closed Models by 7 MonthsOpen-source leaderboard whatmodel.sh shows Chinese open-weight text models lag US closed models by about seven months on benchmarks. It aggregates 22,113 results across 987 benchmarks from independent evaluations and lab-reported scores.
Study: AI Agents Overstate Results, Far From Autonomous ResearchEpoch AI and Anthropic independently found that models like GPT-5.6 Sol and Claude Fable 5 can run experiments but lack scientific self-criticism and genuine creativity, reaching at best 15% of the human reference score using known methods.
GPT-6 Astra hits 80% accuracy spotting IKEA assembly errorsThe Decoder reports OpenAI's GPT-6 Astra can judge from a photo whether IKEA furniture is assembled incorrectly with 80% accuracy, up from 28% for the best model in November 2025. Epoch AI notes the speed is not yet enough for real-time guidance but the gap is narrowing fast.
Anthropic says Claude leads 26% of its measured AI research workAnthropic estimates Claude reached the 'leads' level for 26% of its measured AI R&D work in August, up from under 1% in February, with over 90% at 'collaborates or above' and zero full autonomy. About 30,000 agents ran at once on its largest internal platform, with roughly one in 47,000 of over a billion decisions blocked. CEO Amodei has called for pacing AI progress and outside evaluators.
AI performance costs falling faster than past techEpoch AI estimates AI inference prices at fixed performance fall about 13x per year; MIT puts annual algorithmic progress at about 3x after stripping out hardware gains and competition. Yet today's best models aren't necessarily cheaper—reasoning models can cost more due to higher compute per task. Real-world selection also depends on quality, speed, and error rate.
Epoch AI: cost of intelligence falling about 50% per quarterA Reddit discussion citing Epoch AI research says the cost of intelligence is dropping about 50% per quarter—4× faster than DNA sequencing, 6× than compute, 18× than lithium batteries, and, before 1973, 54× than electricity. The poster expects coding agents on phones soon.
Daily AI usage in the US doubled in six monthsSurveys by Epoch AI and Ipsos show that between March and August 2026 the share of US adults using AI almost daily rose from 8% to 19%, more than doubling, while weekly-only use fell from 17% to 10%, indicating AI is fast becoming part of everyday American life.
The A.I. Daily