Token-based pricing has made the cost of running LLMs a central concern for developers. The new preprint highlights an asymmetry: some frontier providers charge roughly five times more for response tokens than for prompt tokens, even as others offer discounts on prompt tokens through prefilling. That pricing structure creates an incentive for attackers to inflate output length.
According to the abstract, the authors describe trojans that increase response lengths. By embedding such a trojan in a model, an attacker can make every request more expensive for the victim. Because response tokens are the priciest part of a typical API call, even modest length inflation can multiply costs over many requests.
The paper is an arXiv preprint, and the abstract is truncated, so the full attack methodology and evaluation are not yet visible. Still, the core finding is a useful warning: cost management must account not only for model efficiency but also for adversarial manipulation of output length.