Family best-match
Opus, Sonnet, or Haiku indication when signal supports it. Surfaced as a secondary output, not as a verdict.
Model-specific · Claude
Claude AI detector, tuned for Opus, Sonnet, and Haiku.
Claude outputs have different tells than GPT. Longer context, different closing patterns, a specific register. This detector is weighted for Claude, not the generic baseline.
- Claude 2, 3, 3.5, and 4 family-matching
- Opus vs. Sonnet vs. Haiku indicators
- Free, 5,000 characters, no account
About the Claude detector
Claude is Anthropic's model family, and its output has a different fingerprint than GPT's. Detectors built on GPT-only training data are measurably worse at catching Claude output, that's why we run a Claude-specific classifier alongside the generic one.
What Claude output looks like
- Longer-context coherence. Claude's training favors consistency across long passages. Paragraph-to-paragraph argument tracking is tighter than GPT-3.5 and comparable to GPT-4.
- Caveating register. Claude outputs lean toward hedged language ("it's worth noting," "while there are counterarguments") at higher rates than the GPT baseline. This is tunable at the prompt level but survives into casual use.
- Explanatory preamble. Claude tends to restate the question before answering, and restates assumptions more often than GPT. In edited output, the restatement is often deleted, but its traces in sentence structure remain.
- Opus vs. Sonnet vs. Haiku differences. Opus produces longer, more nuanced sentences; Haiku produces shorter, more direct ones. Sonnet sits between. We report the best-match family when signal supports it.
Versions this is tuned for
Claude 2 (legacy), Claude 3 (Opus, Sonnet, Haiku), Claude 3.5 (Sonnet, Haiku), and Claude 4 (Opus, Sonnet, Haiku). The most recent Claude 4 Opus is the hardest to distinguish from human writing in our benchmarks, we report low confidence rather than false precision on that family.
Known failure modes
Claude output given a strong stylistic prompt ("write in the voice of Hunter S. Thompson") is harder to classify, as with any model. Very short passages (under ~150 words for Claude, slightly more than the GPT threshold) do not produce reliable output. Heavy post-editing by a human dilutes the signal, the detector still scores high on the un-edited fragments, but the overall score migrates toward neutral.
Privacy
Pasted text is stored 30 days and deleted. We do not train on user submissions.
Hedged-register weighting
The classifier gives appropriate weight to Claude's caveating patterns instead of reading them as 'just formal prose'.
Long-form stronger
Most AI detectors are weaker on Claude; our Claude-tuned variant closes the gap on passages over 200 words.
Claude-family performance
Tuned to Claude's hedging and list-heavy structure.
0.93
AUC vs Claude 3/3.5/4
Anthropic's hedge phrases are a strong fingerprint.
Sonnet → Opus
Versions covered
Including Haiku for short-form output.
1.4s
Median latency
Across 800–2,400-character pasted essays.
8
Phrase fingerprints
Statistically rare-in-human, common-in-Claude n-grams.
Frequently asked questions
Is Claude harder to detect than ChatGPT?
On paper yes, particularly for Claude 4 Opus. The gap has narrowed across detector families in 2025–26 as training sets have broadened.
Can you tell which Claude version?
Family best-match (Opus/Sonnet/Haiku) is reported when signal supports it. Major-version matching (Claude 2 vs. 3 vs. 4) is more reliable on older versions and degrades on the most recent releases.
Does the detector work on Claude API output vs. Claude.ai?
Yes, the deployment surface doesn't change the text. The score is the same for app and API.
What about Claude Code or agent output?
Agent output often includes code, file paths, and tool-use markers that aren't natural prose. Treat those outputs as a separate problem; the prose-detector applies only to the prose sections.
Ready to check passage?
Free up to 3,000 characters. No account required for a single check.
Run the detector