How we get to a number, and when we refuse to.

Federal law makes health plans publish what they have agreed to pay. Each plan files a machine-readable file. That file lists the rate the plan holds with every in-network provider, for every covered item and service. The files are republished monthly. They are public. No login, contract or vendor relationship is needed to read them. The rule is Transparency in Coverage, 45 CFR 147.212.

That is the entire basis of this product. We hold no claims, no enrollment, no utilization and no member information of any kind. We report what plans have contracted to pay. We do not know, and never claim to know, how often anyone actually used a service.

Publishing the files was the easy part of the rule. Reading them correctly is the product. Each of the following is a real failure mode that a straightforward percentile over the raw data walks directly into.

1. 1

   A percentage is not a dollar

   Rates carry a negotiated_type. Most are dollar amounts, but some are filed as a percentage of another basis, where 0.95 means ninety-five percent, not ninety-five cents. Pooled into a dollar column those rows drag the bottom of the distribution to nonsense. We measured this in the upstream corpus and it is the single largest contamination class we have found. Any cell whose lower quartile or median falls below $5 is refused outright rather than published.

2. 2

   A rate belongs to a contracting entity, not to a doctor

   Rates attach to the entity that signed the contract, which is often a group, a system or a TIN rather than an individual provider. Treating a filed rate as a specific physician's rate misstates who agreed to what.

3. 3

   One carrier brand is many separate plans

   A national brand is a family of separately contracting entities, and they do not pay the same. Two files under one logo are two relationships, not one, and averaging them together invents a rate that nobody actually pays.

4. 4

   An out-of-state plan's file contains in-state providers

   When a member of one Blues plan is treated in another state, the host plan's arrangement can appear in filings for a market that plan does not sell in. Attributing those rows to the local market makes a national default schedule look like a local negotiation. We resolve every payer to a home market and exclude out-of-market host filings from a market's comparison, which is why some markets legitimately return nothing.

One function decides whether a measured cell is fit to show a human being, and nothing reaches a screen without going through it. These are the actual thresholds, imported from the code that enforces them rather than typed into this page.

This is the largest known defect in the Transparency in Coverage files and we would rather you heard it from us. Payers publish a rate for a provider and a billing code even where that provider would never furnish that service. The federal government's own example, in the December 2025 proposed rule, is rates for podiatrists to perform heart surgery. These are called ghost rates.

A peer-reviewed study of 61 insurers found that 91.8% of all published negotiated rates were ghost rates, with the median insurer's file at 84.3%. Restricted to the hundred most common billing codes the share fell, but only to 70.3%. Estimates across the field range from roughly 60% to 96.5%, and there is no agreed definition of what counts, so anyone quoting a single figure as settled is overstating what is known.

The Departments have proposed requiring payers to strip these combinations from the files. That rule is proposed and not final, so nothing about it is in effect and we do not treat it as a fix that has already happened.

Sources: Muhlestein DB, “High prevalence of ghost rates in transparency in coverage data,” Health Affairs Scholar 3(11):qxaf212, 2025. Congressional Research Service report R48570, “Technical Challenges with Private Health Insurance Price Transparency Data,” June 13, 2025. Transparency in Coverage proposed rule CMS-9882-P, 90 FR 60432, December 23, 2025.

A state average hides the market you actually sell in. Prices inside one state vary more than prices between states. So we report at the metro level wherever the filings support it. When a metro cell does not pass the gate, we fall back to the state. The result says so, and it names the metro it fell back from. A broker in Fresno is never shown a California number labelled as theirs.

One limit you should know about how a rate becomes a market. The filings carry no provider address. They name providers by NPI and tax identifier. We work out the metro by joining those identifiers to the national provider registry. That registry can lag where a provider actually practises, and the Congressional Research Service flags this directly. So the metro is a well-founded inference, not something the source file states. A provider who has moved may still be counted in the market they left.

This is the section we would want to read first, so it is not at the bottom in small type.

1. 1

   We have no claims and no utilization

   We know what a service costs under contract. We do not know how often your population uses it. Price without utilization cannot produce a plan spend forecast, and we will not pretend otherwise.

2. 2

   We cannot give you a total cost by site of service

   Medicare publishes two physician fees. One is for an office setting. The other, lower one, is for a facility setting. That lower figure is not what care costs in a hospital. When a service happens in a hospital outpatient department, the facility bills its own payment separately. That facility payment is not in this dataset. So putting the two Medicare figures side by side as a site of care saving shows you something backwards. We publish the comparison only with that caveat attached. We will not publish a total until we hold the facility side.

3. 3

   Filed does not always mean used

   Files can contain rates for services a provider does not perform, and rosters that do not perfectly reflect who is actually contracted. Volume of filings is a measure of how much was filed, not how much care was delivered. We publish the filing count beside every figure so you can judge it yourself.

4. 4

   The corpus has a build date, not a live feed

   Figures are computed from a corpus built August 26, 2026. Plans republish monthly, so a rate negotiated after that date is not in what you are seeing. The date is printed under every number on this site.

5. 5

   We hold no protected health information

   There is no PHI in this product and there is no path by which member data could enter it: we never ingest a source that contains any. This is price data about contracts between plans and providers.

39 services chosen because a commercially insured working-age population actually uses them and because the price dispersion is wide enough to matter to a plan decision. The picker currently offers the 928 largest metropolitan statistical areas, which is what you can look up today. The underlying corpus computes distributions across a far larger set of metros, and we open them as we verify each one rather than listing a headline number you cannot actually query. A market outside the picker resolves to its state and the result says so. Coverage is deeper in dense markets and thinner in small ones, which is a property of how much was filed there, and the filing count on every figure tells you which you are looking at.

Current corpus: Carrier filings published 2026-06-20 to 2026-08-01. 22 of 30 payers carry a publisher date; 8 do not.