The capabilities of AI models are improving faster and enabling the completion of increasingly lengthy tasks, according to several evaluations. Cutting-edge models from US labs are being released more rapidly than ever, even though government interventions have temporarily limited access to two of the most powerful, Claude Fable and GPT-5.6.
Several organizations measure the share of human work that AIs can accomplish. Evaluations by METR and the UK’s AI Security Institute estimate the duration of a human programmer’s work that an AI can perform from a single prompt, i.e., a single instruction. The GDPval test, for its part, compares human experts to AIs, using professional judges. According to these measures, results are progressing at a rate exceeding an exponential curve.
Epoch noted that Opus 4.7, working alone for 14 hours, built a software package representing between 2 and 17 weeks of human engineering work, at a cost of $251 in tokens. A token is a small unit of text that the model reads and produces. In his own experiments, the author indicates that Fable was able to work autonomously for 9 hours on highly complex software projects, which would have taken a team over a week.
The most advanced models are developed by Anthropic, OpenAI and Google, although Google has not released a new model for some time. A second group of models, nearly at the cutting-edge level, is however Chinese. They lag by 6 to 12 months and are open-weight, meaning anyone can use or modify them after their release, unlike proprietary models. Their operation is therefore inexpensive, even if their performance does not always match that announced by their test scores.
Agents Rather Than Chatbots
Until recently, the dominant usage involved asking the AI for a task, checking the result, and then moving to the next step. This mode of collaboration remains useful, but it is giving way to agents, systems that chain actions without constant human intervention. They rely on a software environment, called a harness, which gives them access to tools, and on applications like Claude Code or OpenAI’s Codex.
A joint study by OpenAI and academic economists shows that this transition is accelerating within the company itself. Legal services and human resources have adopted agents at a rate almost identical to that of technical teams. A quarter of OpenAI employees run at least four agents simultaneously each week.
Another study focusing on Claude Code users indicates that software engineers show a success rate comparable to other professions when using the tool for programming tasks. It is not the profession that makes the difference, according to this work, but expertise: the more a person knows their field, the more they succeed with the tool and the more useful the responses they get from Claude for each prompt are.
A Curve Hard to Perceive
An AI plan drafted before winter 2025 described a system capable of a few hours of work, with a fairly high error rate. A few months later, a single prompt can trigger sixteen hours of work or more. This steady progression, however, gives the impression of successive leaps, because humans perceive exponential curves poorly.
The author believes this dynamic also explains the instability surrounding The AI / news. Public authorities have had to urgently modify their policies on cybersecurity issues, and markets have experienced strong fluctuations when AI seemed to threaten certain business models. In his view, these shocks do not reflect an ongoing maturation, but the gap between institutions moving at human speed and a technical capability progressing according to a different logic.
