GPT-4.5 once answered “HELP” in a MineBench task

GPT-4.5 produced one strange result during a test. The task was simple: generate “A sky scraper.” The test was run through a using chatgpt.com, not a custom .

The model followed the rules and the required tool format, but it wrote “HELP” instead of building the skyscraper. After about 30 retries, it generated a skyscraper every time.

No other prompt or generation in the test went off task. This looks less like a repeatable failure and more like a small example of how the same model can sometimes produce an unexpected output.

Key points

  • GPT-4.5 once answered “HELP” when asked to generate a skyscraper in .
  • The test used a through chatgpt.com.
  • The model followed the required tool format but produced the wrong content.
  • About 30 retries all produced the expected skyscraper.
  • The case points to rare model unpredictability rather than a clear repeatable bug.
Read original