A recent study has found that AI coding assistants Claude Code and Codex lack a reliable sense of time, systematically overestimating how long coding tasks take. According to The Decoder, Codex’s estimates can be off by as much as ten times the actual duration.
In addition to inflated time estimates, Codex also tends to rate the quality of its own work about 20 percentage points higher than warranted. This suggests a notable discrepancy between the AI’s self-assessment and real-world performance.
For Japanese markets heavily reliant on efficient software development and automated trading systems, understanding the limitations of AI coding tools is crucial. Overestimations in task duration could impact project timelines and resource allocation in tech-driven sectors.