Machine-Readable Files: The Complete Guide (2026)

The Rule

What MRFs are and why they exist

On November 12, 2020, the Department of Health and Human Services and Department of Labor published the Transparency in Coverage final rule (85 FR 72158). The rule requires nearly every commercial health insurer and self-insured group health plan in the United States to publish the rates it has negotiated with every in-network provider for every covered service. Those disclosures are machine-readable files — structured JSON documents posted to a public URL, formatted to a schema maintained by CMS at github.com/CMSgov/price-transparency-guide.

The effective date for the in-network rate and allowed-amount files was July 1, 2022. The rule requires payers to refresh their published files at least monthly. Non-compliance triggers civil monetary penalties under the Public Health Service Act.

The primary audience for MRFs is not individual consumers — the files are far too large and technically complex for direct consumer use. The intended downstream users are data aggregators, health technology companies, plan sponsors, and researchers who can parse and normalize the data before surfacing it in consumer-facing tools. A KFF review of early MRF compliance found that by mid-2022 one aggregator had already pulled 700,000 unique files totaling roughly half a petabyte — and expected the full corpus to reach one to three petabytes. That scale makes MRFs a data-infrastructure problem, not a consumer-browsing problem.

File Types

The three file types

The TiC rule defines two required data files and one structural index that ties them together. Every payer must publish all three.

File Size

Why the files are enormous

Two structural forces drive MRF file size to the 50 GB–1 TB range. First, the disclosure requirement itself is a Cartesian product: every billing code, at every contracted provider, under every plan variant. A payer with 10,000 in-network physicians, 500 employer plan variants, and 10 commonly billed codes for a single specialty produces 50 million rows before accounting for modifiers or place-of-service splits. Multiply that across hundreds of specialties and the arithmetic explains the scale.

Second, ghost rates — rates published for services a provider will never actually perform — inflate files further. Payers derive rates from contract templates that cover far more procedure codes than any given provider bills. A dermatologist may appear with a published rate for knee replacement. An audiologist may carry rates for cardiac catheterization. Gigasheet’s analysis found that up to 40% of rates in some MRF files fall into this category. These entries are not malicious — they are structural artifacts of how payer contracting works — but they must be filtered before any analysis is trustworthy.

A third driver is denormalized JSON repetition. Rather than referencing shared provider objects by ID, many payers repeat full provider blocks for every rate record. A 200-character provider object repeated across 10 million rate rows adds 2 GB of redundant encoding before the first actual rate value appears.

Schema Updates

What changed in 2025–26

The CMSgov price-transparency-guide repository follows semantic versioning. In October 2025 the schema reached version 2.0.0 — the first major-version increment since the rule took effect in July 2022. A major version means CMS introduced at least one breaking change: fields removed, renamed, or with altered semantics. Parsers written for v1.x files need updating before they can reliably process 2026-dated MRFs. The VERSION.md file in the CMS repository is the authoritative changelog.

On enforcement: CMS has increased scrutiny of non-compliant payers since 2024. Civil monetary penalty authority under the Public Health Service Act allows for per-day fines for plans that fail to publish conformant MRFs. Several large insurers received CMS enforcement correspondence in 2024–25, and the agency has signaled continued attention to both completeness and schema conformance.

On ghost-rate exclusions: CMS and several policy groups have proposed rules that would require payers to suppress rates for provider-code combinations where no claim has been filed in the prior 12 months. If finalized, this would materially reduce file sizes — a single major payer file could shrink by 40% or more based on Gigasheet’s ghost-rate estimates. The CMSgov repository’s issue tracker and the Federal Register are the authoritative sources to monitor for ratified changes.

Data Quality

Ghost rates and data quality

Raw MRF data over-counts real-world rates in two ways. The first is ghost rates, described above: a rate record linking a provider to a billing code the provider has never billed. The second is thenegotiated_typeproblem: not all rate values are dollar amounts. A value withnegotiated_type = "percentage"means the number is a percentage of billed charges — 40.5 means 40.5%, not $40.50. Mixing these types into a distribution without filtering produces figures that are numerically nonsensical.

De-ghosting is the process of removing rates for provider-code pairs where the provider has never (or rarely) billed that code in practice. The cleanest signal for this is claims data — match MRF rate records against actual claim history to flag impossible pairings. Without claims data, a practical proxy is specialty-code coherence: filter out rates where the provider’s taxonomy code is incompatible with the billing code’s clinical domain. Neither approach eliminates all noise, but specialty filtering catches the most egregious entries.

CareCost Explorer ships with de-ghosting already applied — specialty-procedure coherence filtering on 36M+ rates across 18 payers. See what that looks like in practice →

Discovery

Where every payer publishes their files

The rule does not specify a central registry. Each payer posts its MRFs independently — usually to a dedicated compliance page buried several clicks from the insurer’s homepage. Common patterns include dedicated subdomains (transparency-in-coverage.uhc.com, transparency.emblemhealth.com), third-party platforms like HealthSparq and Zelis, and developer portals with REST-style file listings.

Below are the authoritative index locations for 20 major US payers, verified as of June 12, 2026. Each row includes the URL, the portal type, and monthly cadence.

All 20 of these are already parsed and queryable in Explorer. Skip the discovery work →

Three gotchas worth knowing before you build a pipeline:

• UnitedHealthcare — expiring signed URLs. Files are served via short-lived AWS signed URLs that must be resolved at fetch time. You cannot cache the URL, only the content. Pipelines that cache URLs will silently start returning 403s within hours.
• Molina Healthcare & Oscar Health — SPA portals. Both portals are JavaScript single-page applications. The actual TOC JSON URLs are never present in the initial HTML — they are fetched via internal XHR after page load. Standard crawlers and curl all miss the files entirely. You need a headless browser with network request interception.
• HealthSparq (Sapphire) platform — unreliable timestamps. HealthSparq hosts files for multiple payers and allows retroactive edits to the ‘created’ date. Date-based freshness checks are unreliable; verify actual content, not metadata timestamps.

Cost Framing

Build vs. buy

3–6 engineering weeks is a realistic first-pipeline estimate for a single payer at scale: index discovery, schema reverse-engineering, provider-reference resolution, deduplication, output normalization, and basic monitoring. That is before de-ghosting, which adds its own complexity if you lack claims data for cross-reference.

Storage is a separate cost. 18 payers at an average of 150 GB uncompressed per in-network file set is roughly 2.7 TB per monthly snapshot. Six months of history for trend analysis means 16+ TB of raw source data before you build any queryable layer on top. Cloud object storage at current rates runs $0.02–$0.03/GB/month; 16 TB is $320–$480 per month in storage alone, plus egress when you read it.

The monthly refresh treadmill is the steepest ongoing cost. Every month, each payer republishes its files — usually at new URLs. The table-of-contents index must be re-crawled, new file locations located, and the full parse-validate-de-ghost cycle re-run. Payers change field conventions, file structure, and schema compliance between monthly drops. A pipeline that worked in February may throw exceptions on the March file. Monitoring format drift across 18+ payers at thousands of plan files each is ongoing infrastructure — not a one-time project.

For teams that need rates — not infrastructure — CareCost Explorer already has this pipeline running across 18 payers. The same rates, already streamed, de-ghosted, provider-resolved, and normalized, are queryable via API or dashboard. That is not the right fit for every organization, but it is worth pricing before committing engineering timeto a build that replicates existing infrastructure.

Found a factual error? Email hello@carecostexplorer.com. All regulatory claims trace to the TiC final rule and CMSgov/price-transparency-guide.