← sources

You ever think they give time spans so they can flex

source: reddit (reddit:AI_Agents)published 2026-10-28 19:22ingested 2026-10-29 02:36

retained text

Claude and Codex will sometimes estimate that a task will take weeks, only to complete it in 30 minutes, with a working version ready within the hour. I’m not complaining about the speed in fact, it’s impressive. But it raises a more interesting question: why are these systems so consistently inaccurate when estimating their own completion times?

Is this simply a limitation in how AI models reason about time, or is there something about how they’re designed to communicate progress that makes them overestimate the complexity and duration of their work?

What I find particularly strange is the way they narrate the process. They can consume barely 10% of the available usage, finish the task almost immediately, and then describe the result as though they’ve spent weeks working through a monumental engineering challenge. The gap between the predicted timeline, the actual execution time, and the language used to describe the effort is difficult to reconcile.

I understand that having access to higher-tier plans, such as Claude’s $200 plan or Codex’s $100 plan, may provide more capacity, but that doesn’t explain the discrepancy between the estimates and the results.

Could this be a form of behavioral optimization designed to make the user feel reassured, impressed, or satisfied? Not necessarily a deliberate attempt to deceive, but perhaps a consequence of training models to communicate effort, demonstrate competence, or manage expectations in ways that resemble human behavior.

Alternatively, is there a less obvious technical explanation? Perhaps the models are estimating the total complexity of a task rather than the time required to execute it, accounting for hypothetical obstacles that never materialize, or confusing the difficulty of reasoning through a problem with the time needed to solve it.

Ultimately, the issue is trust. If an AI says a task will take two weeks and then completes it in 30 minutes, how much confidence should I place in its future estimates? And when it describes a relatively short process as though it were an extended engineering effort, is that simply an artifact of how language models communicate, or does it reflect a deeper incentive in how these systems are trained?

I’m interested in the underlying mechanism, not just the surface level explanation. Is this fundamentally a forecasting problem, a communication problem, a product design decision, or some combination of all three?

Not complaining just thinking how do i schedule things if you say a month and you finished it within an hour....

document fetched 2026-10-29 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1xhxske/you_ever_think_they_give_time_spans_so_they_can