Epoch (23 Oct): six frontier models try 11 of its research tasks unaided; none reliably meets its quality bar
Epoch AI had six leading models attempt 11 realistic tasks from its own work, one try each with no help: Claude Fable 5.1 and GPT-6 Astra tied for first, but none reliably met Epoch's standard, doing well on coding and data analysis and poorly on experiment design and interpretation. Its index puts Gemini 4 Argon at 166, behind Opus 5.5 and GPT-6 Astra at 169.
Why it matters. A careful test by a research group finds that today's best AI models can do its well-defined tasks but cannot yet reliably do its open-ended research work unaided. If this holds more widely, replacing skilled workers outright might take longer than the recent wave of job-cut announcements suggests, even as routine parts of their work are automated.
Next. The next published measures of how long and how open-ended a task AI can complete on its own.
Sources: epochai.substack.com · Point 1 of our view