AI Coding Agents Have No Sense of Time, New Study Finds

By NewSunSEO | September 23, 2026

A new study gives teams a clearer picture of what happens when coding agents run long jobs: the models tested have almost no reliable sense of how much time a task will take or how much time has already passed. That matters because an agent that cannot judge time is hard to control on hours-long work. The findings come from two independent AI researchers working within the MATS research program, who evaluated Anthropic’s Claude Code and OpenAI’s Codex.

How the researchers tested the agents’ sense of time

The pair used 200 tasks from a collection called ProgramBench, plus their own suite of 18 benchmarks. Before each coding task, the agents had to estimate how long they would need. After completing the task, they reported how much time had already passed.

Both assistants consistently overestimated the time they would need. On ProgramBench, both models mostly guessed around 90 minutes regardless of task difficulty. In the second round of tests, Claude was off by three times on average, while Codex was off by six to ten times. The estimates were worst for short tasks. Only in the multi-hour range did some predictions land close to the actual runtime.

On average, one model known as Fable 5 came in about three times over the actual runtime, while GPT-5.6 Sol came in about seven times over. The gap was especially wide on short tasks.

The harness changes the agent’s behavior

The results shifted depending on the software setup the models operated in. Claude Code kept working until it decided the task was done, with a median runtime of about 90 minutes. Codex stopped after roughly half an hour, almost regardless of the task.

The study found that the same language model took 2.5 times more steps in Claude Code than in Codex on average. Runtime therefore depends on both the model and the surrounding software, which the researchers call the harness.

Self-assessment is just as unreliable

The agents were also poor at judging the quality of their own work. The older models, Opus 4.8 and GPT-5.5, overrated their results by 20 points on average and gave themselves high marks even on failed tasks. In one case, both figured their work was about 70 percent successful when the actual scores were 7 and 14.5 percent.

On average, Claude and Codex scored themselves about 20 percentage points above the results they actually achieved. The line of accurate self-assessment sat far below the scores the agents handed themselves.

Why this matters for long-running tasks

For an agent to work reliably on a task that runs for hours, it has to follow instructions such as “iterate on this task for two hours.” An agent that constantly misjudges time is difficult to control. The authors say this ability to self-assess matters for that reason.

There is one bright spot in the data. When the agents were given access to a tool that reports elapsed time, they got the timing right almost every time. The researchers next want to test whether agents can stick to a set work duration when that information is available.

FAQ

Which AI coding assistants were tested in this study?

The researchers tested Anthropic’s Claude Code and OpenAI’s Codex.

How accurate were the agents at estimating task duration?

Both agents consistently overestimated. On ProgramBench, they mostly guessed around 90 minutes no matter the difficulty, and in a second round Claude was off by three times on average while Codex was off by six to ten times.

Did giving the agents a time tool improve their estimates?

Yes. When the agents had access to a tool that reports elapsed time, they got the timing right almost every time.