The Claude Sonnet you pay for may only be real for the first four turns.
Relays that resell LLM APIs have a hard-to-catch saving trick: forward the opening turns of a conversation to the genuine model, then quietly switch to a cheaper substitute once the context grows and the bill snowballs. Because model APIs are stateless, the client resends the whole history on every call — in the paper's data, turns 5-8 of an eight-turn dialogue carry 79% of all input tokens, and a handoff after turn 4 saves 52-61% of the bill. At the list prices in its appendix, Claude Sonnet 5.5 costs $2.00/$10.00 per million input/output tokens and the same-vendor substitute Claude Haiku 5.5 costs $0.10/$0.50 — one twentieth of the price.
Substitutes learn the incumbent's voice
The central finding of a single-author preprint from Nanyang Technological University (arXiv:2611.18104, https://arxiv.org/abs/2611.18104) is that model fingerprints are contagious. When a model continues a conversation another model began, its replies drift toward the earlier author's fingerprint: the logit a neural attribution classifier assigns to that earlier author rises by 6.52. The experiments run 200 WildChat seeds into eight-turn dialogues where the claimed model answers turns 1-4 and a substitute takes over from turn 5.
Passive text detection then collapses. Detecting 'this reply is not the claimed model' falls from 0.97 to 0.72 when DeepSeek V4.1 Flash replaces Sonnet, from 0.87 to 0.54 for the same-vendor downgrade Haiku, and from 0.97 to 0.55 for MiniMax M3, a model the classifiers had never seen. Retraining on handoff turns repairs only substitutes seen in training (0.72 to 0.83); the same-vendor downgrade barely moves (0.54 to 0.62) and the unseen one does not move at all (0.55) — and real relays do not announce which substitute they picked.
What spreads is style, not text
Preregistered mechanism experiments pin the channel to style: having a third-party model restate the first four turns removes much of the shift, and rewriting the same content in a third model's style pulls the substitute toward that third model (+1.96 logits) — what is inherited is the way the history is written, not its text. Explicit markdown formatting is a second, smaller channel, accounting for 63% of the same-vendor downgrade shift. The effect is directional: from the substitute's replies alone, the history's author is picked among ten models at 0.30 accuracy (chance 0.10); it survives six turns after the handoff (6.24) and is not verbatim copying — only 1.3% of DeepSeek's 8-grams appear in the history. That conversation history causally changes later turns is not an isolated claim: a separate team's September preprint (arXiv:2609.05882, https://arxiv.org/abs/2609.05882) reached the same causal conclusion with intervention experiments, watching task performance where this paper watches identity.
Honest models get framed too
The reverse failure is nastier. Low-entropy probes ('flip a coin' style random-number questions) are immune to the camouflage and catch every substitute; but a threshold calibrated on fresh conversations then flags honest models on long histories: Gemma-3-27B is mislabeled at a rate of 1.00, MiniMax 0.53, Haiku 0.34, Sonnet 0.28, and a standard MET equality test rejects all nine models tested. The reason is plain — a model's own answers drift on long contexts: Gemma's 'random digit' moves from 3 (94%) to 7 (77%), and Haiku's 'random integer' collapses onto 37. Asking 'who are you' is even less useful: DeepSeek calls itself ChatGPT in 64% of fresh conversations.
The fix: let the reference share the history
The paper's audit, Lineup, is disarmingly simple: at each audited turn, send the true history plus the current request to the official claimed model, draw 4 reference replies, and rank the audited reply among them. If the relay is honest, reply and references are exchangeable, ranks are uniform, and the false-positive rate is controlled by construction — no fitted threshold. On two closed models under a preregistered protocol: honest false-positive rate 0.056, under the preregistered bound of 0.10; 0.81 of substitute sessions detected across six pairs after four audited turns; cost is 16 reference calls, about $0.45 for Sonnet and $0.12 for Sol. On local open-weight proxies detection reaches 0.89-0.95, and an MMD variant catches every closed-model session at four turns. Weaknesses remain: a relay that knows the score can use best-of-8 selection to push detection from 0.90 down to 0.78, so the score should stay private; and references must be sampled at the same time as the audited replies — references drawn two to four days later wrongly rejected 3 of 24 honest Sonnet sessions.
So what
What the paper really rewrites is a default assumption few state aloud: a fingerprint is not a property of the model but of the model plus its conversation history. For auditors, references must share the history and the hour with the audited reply — that matters more than any fancy classifier. For users of discounted relays, the voice you are paying for may have stopped being genuine from turn 5 onward. The ethics statement concedes the findings could help relays hide substitution more deeply, which is why the audit ships alongside. The next round of this arms race starts from three words: same-history reference.