StealthGPT claims 89% Pangram bypass rate in own benchmark, separately reports 18% factual-consistency failure rate

A humanizer moves text away from a classifier's decision boundary, and the only lever it has is the text. Word choice, sentence length, punctuation habits, connective structure: those are the features detectors read, and they are also where the facts sit. Rewrite for surface statistics and you are rewriting the tokens that carry numbers, names, hedges and attribution. So factual drift is part of the bill for the edit.

StealthGPT says this out loud. Gherman wrote that a rewrite can sound natural and evade detection while changing claims, numbers or meaning ^[7]. Super's answer is a multistage pipeline that inspects structure, factual consistency, completeness, naturalness, stance, sentence integrity and omissions, then checks the output again after making fixes ^[8]. That is the right shape for the problem: generate, verify, repair, verify again. The 82 percent pass rate is what survives it ^[6].

The report publishes two margins and not the cell where they meet. Eighty-nine of 100 rewrites read as human to Pangram V4 ^[3], and 82 of 100 passed the factual-consistency check ^[6]. If those outcomes were independent, 0.89 x 0.82 gives about 73 rewrites in 100 clearing both ^[16]. Independence is an assumption I cannot test from what was published, and I would expect the correlation to run the wrong way, because heavier rewriting is what buys evasion and what costs accuracy. StealthGPT is currently the only party holding the overlap.

For 89 percent to describe someone else's workload, the sample would have to match that workload's topic mix, document lengths and source model, the selection would have to be independent of whatever Super was tuned against, and the detector on the other end would have to be the same build. Runtimewire's account of the test is explicit that it does not establish an 89 percent bypass rate across other datasets, writing styles, document lengths or future Pangram versions, and that detector performance moves with topic, language, editing history and the model that generated the original text ^[12]. The build under test shipped on July 29, 41 days before the launch post ^[17].

Pangram's own V4 benchmark claims it caught AI involvement in 98.83 percent of samples processed by 13 commercial humanizers, which leaves roughly 1.17 percent through on Pangram's data ^[9][18]. Both figures can be true at once, since neither measures the other's dataset, and no neutral head-to-head on an agreed test set has been published ^[11]. The repository dates the Super run September 7 while the announcement says the samples were tested September 6 ^[5]. A one-day gap in a self-published test proves nothing except that nobody is auditing the lab notebook.

The funding is the part that sets tempo. Pangram raised $9 million alongside V4, while Gherman describes StealthGPT as bootstrapped on personal savings and friends-and-family capital ^[10]. Gherman started the company in February 2023 ^[15] and, by his own account, shipped Super because Pangram V4 broke the previous models ^[1]. Runtimewire's read of the commercial logic is that each detector update both threatens the core product and hands existing customers a reason to pay for the replacement ^[19]. StealthGPT's API page reports more than 700 API users, over 350 businesses and more than 50 million generated words as of September 9, 2026, all self-reported ^[13].

Until the overlap is published, about 73 in 100 is the more honest planning figure, and it is one I derived from StealthGPT's two margins rather than one the company stands behind.