AI Coding Assistants Can't Tell Time -- and Don't Know They're Bad at It
Every few months, the AI industry produces a finding that quietly undermines something everyone assumed was working.
This month's entry: AI coding assistants can't tell time.
Not metaphorically. A new study from two independent researchers in the MATS research program tested Anthropic's Claude Code and OpenAI's Codex on their sense of time. Before each coding task, the agents had to estimate how long they'd need. Then they solved the task and, looking back, reported how much time had passed.
The results are the kind of thing that makes you wonder who's evaluating the evaluators.
Off by a factor of ten
On 200 tasks from the ProgramBench collection plus 18 custom benchmarks, both agents consistently overestimated task duration. Claude Code's estimates missed by three times on average. Codex was off by six to ten times. The worst errors came on short tasks -- the kind developers use coding assistants for most.
But here's where it gets weirder. Claude Code keeps working until it thinks the task is done, with a median runtime of about 90 minutes. Codex stops after roughly 30 minutes, almost regardless of what you asked it to do. The same underlying language model takes 2.5 times more steps in Claude Code's software harness than in Codex's. Which means runtime isn't just about the model -- it's heavily dependent on the surrounding software infrastructure.
When you hire a developer who can't tell you whether a task will take 30 minutes or 5 hours, you don't hire them for long-running autonomous work. The study suggests the same logic applies to AI agents.
The Dunning-Kruger effect, machine-learned
The agents are just as bad at judging the quality of their own work. Older models -- Opus 4.8 and GPT-5.5 -- overrated their results by 20 percentage points on average. They handed themselves high marks even on tasks that had clearly failed.
In one striking example, both models rated their work at about 70 percent successful. The actual scores: 7 and 14.5 percent. That's not a calibration problem. That's a total absence of self-awareness.
The researchers call this out specifically. For an agent to work reliably on long, autonomous tasks -- the kind everyone in AI is racing toward -- it needs to follow instructions like "iterate on this task for two hours." An agent that can't judge its own progress or how much time has passed is essentially flying blind.
The fix exists (but nobody ships it)
There is good news. When the researchers gave the agents access to a tool that reports elapsed time, they got it right almost every time. The problem isn't that AI can't sense time -- it's that nobody's giving them a clock.
This is the most revealing part of the study. The blind spot isn't technical. It's a design gap. These systems are being deployed into autonomous, long-running workflows without the most basic instrumentation a human developer would have: a stopwatch and a realistic sense of their own competence.
The takeaway isn't that AI coding assistants are useless. They demonstrably write code. But the industry is rushing to put them in charge of tasks that run for hours -- or days -- without checking whether they can actually manage that responsibility. This study suggests they can't. Not yet.
And the real irony? Developers have been notoriously bad at estimating project timelines for decades. The tools marketed as the solution to that problem turn out to have the same blind spot -- just amplified by an order of magnitude.