The AI Model You Paid For May Not Be the One You Got

An API call for 'claude-fable-5' may return a completion from 'claude-opus-4-8' due to a classifier routing sensitive requests. Cursor's Router dispatches queries to different models based on context, claiming 30–50% cost savings. OpenRouter warns that some providers serve quantized weights, but logs do not reflect this. Model identity is fracturing along three axes: substitution (different architecture), degradation (reduced precision), and drift (silent weight updates). This raises contract law issues: if a buyer specified a model name, substitution is breach; if they bargained for capability, a definition of 'frontier quality' is needed. Under FTC deception doctrine, disclosure varies: Anthropic notifies and returns the served model, Cursor publishes rules but not per-task assignment, OpenRouter buries quantization info in docs. For evidence, Federal Rule of Evidence 901(b)(9) requires naming the system, but routing breaks the chain of custody. The author argues for attestable model identity: a signed assertion with served model, weights hash, precision, and system prompt hash verifiable without trusting the seller.