The twilight of the chatbots

The newsletter One Useful Thing published a piece titled "The twilight of the chatbots" arguing that AI progress is accelerating. The post states that better models from leading American AI labs have been releasing more quickly than ever, adding that government interventions temporarily stopped access to two of the most powerful models, Claude Fable and GPT-5.6. It says the evidence points to accelerating capability gains as well, while noting the frontier stays jagged and AIs remain weak in many areas.

The post cites several assessments measuring how much human work AIs can do. METR and the UK's official government AI Security Institute estimate the amount of human programmer hours of effort an AI can do with a single prompt, and GDPval compares human experts in many fields to AI performance using professional judges. The post says all are increasing at a better than exponential rate.

It also cites Epoch, which recently found Opus 4.7, working on its own for 14 hours, was able to build a software package that would take 2-17 weeks of human engineering work, at a cost of $251 in tokens. The author's own experiments found Fable worked autonomously for 9 hours on complex software projects the author says would have taken a team well over a week.

The post distinguishes frontier models, made by three American companies — Anthropic, OpenAI, and Google — noting it has been a while since Google released a new model. It describes a second set of near-frontier models from China that typically lag 6-12 months behind, are open weights, and are cheap to operate. It says these are on their own exponential curve, behind closed US models, as shown in a graph of performance on a test called AA-Briefcase that simulates a complex multi-week consulting engagement. It adds that other countries produce open weights models, but none are near the frontier.