A preprint on arXiv introduces API-TAIL, a method designed to detect cases where some requests are quietly switched to cheaper models. The concern is that commercial language-model APIs return only text, so in theory, a provider could route some requests to cheaper alternatives while charging the rate for the original model. Detection is difficult because legitimate providers also adjust temperature, truncate sampling, add frequency and presence penalties, and change default system prompts—all of which can shift the output distribution. Auditing therefore needs to account for a range of legitimate configurations, not just a single baseline. API-TAIL aligns an effective temperature across a batch, centers the standardized surprise scores, and then checks whether the number of remaining tail responses exceeds a threshold calibrated across legitimate configurations. The authors say it controls false positives on real commercial APIs and can detect low replacement rates with a small query budget, outperforming ranking- and MMD-based audits. This is a preprint and has not been peer-reviewed.

$BTC