ModelReceipts: proof that the model you paid for is the model that answered
The problem
Teams buy frontier models through resellers, gateways, and marketplaces — OpenRouter, relays, regional providers — and then have no way to verify what they got. The model string in the config says GPT-5 or Claude Opus; the backend may quietly serve a cheaper model, a heavy quantization, or a previous generation. A 2026 survey of 17 third-party wrappers found 45.8% had endpoint-level identity failures, including wholesale swaps like GPT-5 to GLM-4-9B, and OpenRouter users are already asking in public whether providers quietly serve cheaper models behind the name they pay for.
The idea
Why now
Three things converged this month: Anthropic's FlashProve (2 Nov) made model identity cryptographically provable with zkSNARKs; the AgentProv paper (arXiv 2609.00052) showed tool-call distributions fingerprint a backend even when the text channel is discarded; and OpenRouter's Auto Router (6 Nov) made multi-provider routing a default — which means the substitution surface just grew, because every routed call is a call whose backend the buyer cannot see.
What it combines
FlashProve's cryptographic proof of model identity + OpenRouter Auto Router's multi-provider routing + Sarvam Prism's 44 ms decision scoring + GPUWorks' serving telemetry. Proofs without routing context have nothing to audit, routing without proofs is blind trust, and cheap per-call verdicts without either are just another log line — together they are a continuous audit: every call fingerprinted, sampled calls proven, and every invoice reconcilable.
MVP
Weekend MVP: a proxy that sits in front of one provider, runs calibrated text probes against a known reference for one model pair, and emits a daily pass/fail identity report with the statistical confidence attached. Skip: zk proofs, tool-call fingerprinting, multi-provider support, and invoice reconciliation.
How it makes money
The moat
Providers cannot credibly audit themselves — a reseller that offers its own "proof" is asking buyers to trust the party being checked, so the independent auditor is structurally necessary. Observability incumbents would have to rebuild around statistical identity testing and zero-knowledge proofs, a different architecture from request logging, and Helicone itself is in maintenance mode after its March 2026 acquisition, leaving the category leader stalled.
Distribution
B2B via API gateways and routers — an add-on audit inside OpenRouter, Portkey, and LiteLLM-style gateways, so every team routing through them gets receipts by default. Direct sale to platform-engineering and FinOps teams buying AI APIs through resellers.
Why it wins
Helicone, Langfuse, and LangSmith log requests, costs, and traces — none of them can tell you whether the backend was the model you bought. Fingerprinting papers exist as research, not as a product. ModelReceipts is the only service whose deliverable is an identity verdict per provider, not a prettier cost dashboard.
Risks
Biggest risk: false accusations — a fingerprint test that flags an honest provider destroys trust in the auditor faster than a missed substitution does. The MVP de-risks it by publishing the false-positive rate on known-honest controls alongside every verdict, and by starting with detection (flag for review) rather than automated billing disputes.
Build it with
- Anthropic FlashProve: cryptographic proof of which AI model actually answeredCryptographic proof of which model answered — the receipt format for providers that cooperate.
- OpenRouter Auto Router: next-gen model routing with a dedicated coding routerThe multi-provider routing layer where substitution happens — the audit sits here.
- Sarvam Prism: open-weight decision model that answers in 44 milliseconds44 ms per-call verdicts, so fingerprint sampling runs continuously without slowing traffic.
Repo to start from
modelreceipts — a proxy that fingerprints the model behind any AI API endpoint and emits a daily identity verdict with statistical confidence.
Evidence
- AgentProv: auditing agentic LLM API providers via tool-use policy probes
- Which model is your API actually serving? Four ways it has been checked in 2026
- How to check whether your Claude API relay is serving the real model
- Helicone pricing: Pro $79/mo for proxy observability
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.