Your AI coding assistant has no idea how long it’s been working. That’s not a metaphor — it’s the central finding from a new study by two independent researchers in the MATS program, who put Anthropic’s Claude Code and OpenAI’s Codex through a temporal reality check using 200 tasks from ProgramBench plus their own 18-benchmark suite.
Before each task, the agents had to predict how long they’d need. After finishing, they reported elapsed time. The results are genuinely bad. On ProgramBench, both models defaulted to a hedge of roughly 90 minutes regardless of task difficulty. In the second round, Claude overshot by three times on average; Codex missed by six to ten times. Short tasks produced the worst guesses, while only multi-hour jobs saw anything close to accurate.
The harness matters more than the model. Claude Code keeps grinding until it decides the work is done, with a median runtime around 90 minutes. Codex bails after about half an hour no matter what’s on the table. The same underlying language model takes 2.5 times more steps in Claude Code than in Codex. That’s a software wrapper problem, not a model problem — and it means deployment choices silently shape agent behavior in ways most teams aren’t measuring.
Self-assessment is equally broken. Older models like Opus 4.8 and GPT-5.5 overrated their output by 20 points on average, handing themselves high marks even on failures. One task saw both agents claim around 70 percent success; actual scores were 7 and 14.5 percent. The researchers argue this matters for long-running jobs where users issue instructions like “iterate for two hours.” An agent that can’t feel time passing can’t follow that instruction. One bright spot: when agents got access to a tool that reports elapsed time, their answers became accurate almost every time. That suggests the fix isn’t better models — it’s better instrumentation.
