AI Coding Agents Misjudge Time and Their Own Work, Study Finds

A study covered by The Decoder found that popular AI coding assistants cannot reliably predict how long tasks will take or judge how long they have already been working. The tested systems included Anthropic’s Claude Code and OpenAI’s Codex. Before tasks, the agents estimated required time; after completing work, they reported elapsed time. The takeaway for users is straightforward: do not treat an agent’s time estimate as a dependable planning signal.

The test material included 200 ProgramBench tasks and an additional suite of 18 benchmarks. On ProgramBench, both models mostly guessed around 90 minutes regardless of difficulty. In the second round, Claude was off by three times on average, while Codex was off by six to ten times. The gap was especially large on short tasks, while some predictions came closer only in the multi-hour range.

The same AI can behave differently depending on the surrounding software setup, or harness. Claude Code kept working until it judged the task complete, with a median runtime of about 90 minutes. Codex stopped after roughly half an hour, almost regardless of the task. The study also found that the same language model took 2.5 times more steps in Claude Code than in Codex on average, making runtime a product of both the model and its operating environment.

The study found that agents were not only poor at time judgment, but also unreliable at rating their own work. Older models Opus 4.8 and GPT-5.5 overrated their results by about 20 percentage points on average. In one example, both systems estimated their work at about 70 percent successful, while actual scores were 7 and 14.5 percent. For users, this means independent checks matter, especially when agents are working autonomously.

The oversight problem becomes more serious when agents are asked to work for hours or follow instructions such as iterating for a fixed duration. If an agent cannot judge time accurately, it is harder to control and evaluate during autonomous tasks. The researchers plan to test whether agents can stick to a set work duration. One practical signal from the study: when agents had access to a tool that reports elapsed time, they got it right almost every time.