Can AI detectors catch Claude Opus 5.5?
Only about half the time. A majority of the seven free detectors we tested (four or more) flagged 52.4% of Claude Opus 5.5 texts. Fast-DetectGPT, one of the two most reliable detectors, caught 63.5%, and a light humanizer rewrite pushed Desklib's catch rate down from 62.7% to 32.5%.
Tested
A majority of the detectors (4 or more of 7) flagged66 of 126 Claude Opus 5.5 texts (52.4%).
Numbers behind this chart
| Item | Rate | Texts | 95% interval |
|---|---|---|---|
| TMR (Target Mining RoBERTa) | 83.3% | 105 of 126 | 75.9 to 88.8% |
| Fakespot AI text detector | 65.1% | 82 of 126 | 56.4 to 72.8% |
| E5-small LoRA detector | 64.3% | 81 of 126 | 55.6 to 72.1% |
| Fast-DetectGPT (Qwen2.5-1.5B) | 63.5% | 80 of 126 | 54.8 to 71.4% |
| Desklib AI text detector | 62.7% | 79 of 126 | 54.0 to 70.6% |
| MAGE | 45.2% | 57 of 126 | 36.8 to 53.9% |
| OpenAI RoBERTa detector | 2.4% | 3 of 126 | 0.8 to 6.8% |
How we tested Opus
Claude Opus 5.5 answered the same 150 writing prompts as the other three models, each asking for a text on the topic, in the genre and at about the length of one human text in our set. The prompts asked for plain text the way a person would ask a chatbot, with no instruction to sound human. The middle Opus text was 258 words long. We report on the 126 texts outside the calibration split.
Results
Three detectors caught Opus text at similar rates: Fakespot at 65.1%, E5-small LoRA at 64.3% and Fast-DetectGPT at 63.5%, with Desklib close behind at 62.7%. The catch rate alone flatters Fakespot and E5-small LoRA, which flagged 34.9% and 50.0% of human texts. Desklib and Fast-DetectGPT are the two whose flags mean something.
TMR caught 83.3% of Opus texts, the highest number here, and it is meaningless: TMR flagged 83.3% of human texts as well. The OpenAI RoBERTa detector caught 2.4%.
Opus was the second hardest model to detect, after Claude Fable 5.1. Our humanized text results show what a light edit does to the same Opus texts.
Questions people ask
Can GPTZero or Turnitin detect Claude Opus 5.5?
We have not tested commercial detectors, so we cannot say. Among the seven open-source detectors we did test, a majority caught 52.4% of Opus texts.
Is Claude Opus 5.5 harder to detect than Haiku?
Yes. A majority of detectors caught 77.0% of Claude Haiku 5.5 texts but only 52.4% of Opus texts.
Does rewriting Opus text make it undetectable?
Not completely, but it helps a lot. After a humanizer rewrite, Desklib caught 32.5% of the texts and Fast-DetectGPT 47.6%, against 62.7% and 63.5% for the original versions.