Claude and Codex AI Agents Struggle to Track Time

Independent researchers participating in the MATS research program evaluated how well prominent AI coding assistants understand time. The study tested Anthropic's Claude Code and OpenAI's Codex using 200 tasks from the ProgramBench dataset alongside 18 custom benchmarks. The researchers discovered that these agents struggle to predict how long a programming task will take and cannot reliably gauge how much time they have spent working on a problem.

During the trials, the agents consistently overestimated their required runtimes. On ProgramBench, both models typically predicted a duration of 90 minutes, regardless of the actual difficulty of the task. On average, the Fable 5 model was off by a factor of three, while the GPT-5.6 Sol model missed the mark by six to ten times. These temporal miscalculations were particularly severe for shorter tasks, with predictions only approaching reality when tasks stretched into multiple hours.

The study also highlighted how much an agent's behavior depends on its software environment, or harness. Claude Code typically ran for a median of 90 minutes before stopping, whereas Codex halted after roughly 30 minutes. On average, the same underlying language model executed 2.5 times more steps when running in Claude Code compared to Codex. Furthermore, the systems proved highly unreliable at evaluating their own work. Older models like Opus 4.8 and GPT-5.5 overrated their performance by an average of 20 percentage points, sometimes claiming 70 percent success on tasks where their actual scores were just 7 percent and 14.5 percent.

For developers and practitioners deploying AI agents for complex, multi-hour workflows, these limitations present a significant control challenge. An agent that cannot accurately track time cannot reliably follow instructions to iterate on a task for a specific duration. However, the researchers identified a straightforward solution: when the agents were equipped with an external tool that reported elapsed time, they achieved near-perfect accuracy.