The Top Repeated Head-CT Rate Is $119 in One of This Insurer's Networks and $380 in Another. The Ranking Underneath Mostly Holds.

An insurer does not publish one price list. In a single month BlueCross BlueShield of Tennessee publishes fifteen separate home-network files — one per network product — alongside hundreds of employer-specific index files, all under the same federally required transparency rule.

We parsed fourteen of those fifteen for the July 27, 2026 vintage — 2,462,194 provider-code rates in all — and asked a question that only becomes askable once you have more than one file from the same insurer: when the same practice appears in two of a payer's own networks, does its price position travel with it?

The answer matters because of what happens to the labels first.

The same code, the same insurer, the same month, four different top prices

Take the four largest of those networks. Two carry roughly 57,000 providers each; two are smaller, at about 32,000 and 27,000. For a routine CT of the head, here is the highest of the five most frequently repeated prices in each file — the number a benchmark keyed to "the top of this payer's schedule" would pick up. The provider count is the network's full roster; the number of those actually priced on a head CT is smaller (46,195 in panel A, 17,247 in panel D):

That is a 3.2-fold range in a number produced by one company, in one state, in one month.

And it does not move in a consistent direction. You cannot summarize it as "panel C is the cheap network." Run the same comparison on a session of therapeutic exercise and C flips to the most expensive of the four, tied with panel D:

Panel C carries the lowest top repeated rate of the four on the scan and the injection, and the highest on the therapy, where it ties panel D to the cent. The multiples against each network's own median run from 1.09× to 3.85× depending on which of the insurer's products you happen to open.

There is a second, quieter finding in that table. Panels A and B are not two negotiations. Thirteen of the seventeen codes we examined carry the identical top repeated rate in both files — the same number, to the cent, on the CT, the MRI, the ultrasound, the X-ray, the therapy session, the blood test. The four that differ do so by between 2.3% and 6.4%, and two of those four — the level-3 and level-4 office visits — differ by a multiplier that agrees to five decimal places, 1.06368 and 1.06373. Two of this insurer's separately published networks are substantially one price list emitted into two files, with a handful of lines nudged.

So the tier label is not a stable object. It changes by threefold between products of the same company, it reorders between services, and two files that look independent are one schedule published under two names.

Then what, if anything, does travel?

The provider.

We took every practice that prices a given code in both of two networks and compared its rank in one against its rank in the other. Three progressively harder versions of that comparison, because the easy version is misleading and we want to show why:

Read that top to bottom. Each step removes a way the number could be true for an uninteresting reason.

The first cut removes providers sitting on the payer's standard fee schedules. We have published before that this insurer assigns most providers onto proportional rate tables; for anyone on one of those, rank stability is arithmetic, not a discovered fact.

The second cut is the one that matters, and it was the strongest argument against this article. Off-schedule does not mean the rate differs across networks. A practice can be individually priced in both files and carry the identical number in both — in which case its rank agrees perfectly and nothing has been learned about position. That happens a lot. Across all providers, 48.3% of the cross-network cells between panels A and C are byte-identical, and 54.2% between A and D. Even inside the off-schedule population it is still 31.6% and 29.6%. Roughly three off-schedule providers in ten were inflating the correlation for free.

Remove them, and the correlation drops from 0.77 to 0.60 (128,179 provider-code pairs) and from 0.78 to 0.63 (84,549 pairs). Attenuated, not collapsed. That is the honest number, and it is a better sentence than 0.77, because it is what is left after the competing explanation was tested and priced rather than assumed away.

For scale: shuffle the rankings at random within that same discordant population and re-run, and the correlation averages about zero — across three seeds and both comparisons the average lands between −0.006 and +0.006. Per code, the real figures run 0.47 to 0.78 for panel C and 0.49 to 0.79 for panel D. There is no single code carrying the result.

The objection after that one

There is one more way 0.60 could mean something dull: discordant is not the same as individually priced. Two providers could sit on two different standard schedules — different numbers in each file, but still assigned rather than negotiated — and the rank correlation would be schedule-driven one layer further down. Our own earlier work is what raises this: we published that this insurer's imaging tiers sit at an exact 4.23× ratio to one another, and a mechanism that clean would leave fingerprints.

It would leave them in the ratios. If these providers were being moved between a handful of schedules, the ratio of their panel-C rate to their panel-A rate would cluster on a few values.

It doesn't. Among the 128,179 discordant pairs there are 5,638 distinct rate ratios; the single most common accounts for 4.6%, and the twenty most common together cover only 30.5%. Panel D is the same picture: 4,095 distinct ratios, most common 5.1%, top twenty 33.7%. A proportional-schedule mechanism produces a handful of ratios covering most of the population. This is the opposite of that. These practices are priced individually in both files.

(One incidental number worth having: the median ratio is about 0.94. A given practice's rate in the two smaller networks typically runs around 6% below its rate in panel A.)

A statistic we computed and threw away

This piece exists in its current form because an earlier version of it was wrong, and the way it was wrong is worth more to a reader than the finding.

We had a striking number, and it was wrong. Within one network, the top imaging tier, the top physical-therapy tier, and the top injection tier are held by almost exactly the same providers — 3,058 practices hold all three. Against the wrong null — one in which a payer assigns each line's top tier independently — you would expect about 15, and fifteen-versus-three-thousand is the kind of ratio that gets an article written about it. We wrote that article. Then we killed it, because the null was wrong and the right one predicts 3,058.8. This is the piece that replaced it.

We should have known on our own evidence. This payer does not assign tiers line by line. It assigns whole fee schedules — we published that ourselves. Under a null that holds schedule assignment fixed and shuffles everything else, the expected number of providers holding all three tiers is 3,058.8, with a standard deviation of 0.8. The file contains 3,058.

The prediction was right to within one provider. There was no finding. The arithmetic had been correct the entire time; the comparison was against an assumption nobody in this market should hold.

That is not an unusual failure, and it is the reason to be careful with any benchmark that tells you how unusual your rates are. The number is almost never wrong. What it is being compared against usually is. A statistic is only as good as the alternative it rules out, and "these prices were set independently" is an alternative that this market, on its own published evidence, has already ruled out for us.

What we did not measure

Labs are excluded from every correlation and rate-identity figure above; the one comparison that counts them — thirteen of seventeen codes — says seventeen in the figure itself. The exclusion is not cosmetic. A complete blood count is priced here as an assigned commodity — nearly all providers sit on a handful of rates — so the off-schedule, rate-differs population for that code collapses to 114 providers out of 23,950, about half a percent. On that sliver the correlation is −0.86, and it is the same 114 practices in both comparisons. That is noise on an extreme tail, not a finding, and pooling both lab codes back in would drag the headline figures to 0.54 and 0.56 and inject a nonsense negative. We reported it here and left it out of the number. We have published separately that a provider's lab position is unrelated to its position on physician-delivered work; this is that result again, from a different direction.

One methodological limit belongs here rather than in a footnote. "Off-schedule" is defined as not among a code's five most frequently repeated rates in either file. If this payer runs a sixth or seventh standard tier, some assigned providers leak into what we are calling individually priced. The ratio dispersion above bounds that leak — 5,638 distinct ratios is not what a handful of hidden tiers produces — but it does not eliminate it, and a tighter definition would likely move 0.60 and 0.63 somewhat.

We also did not test whether any of this holds for a second insurer. Everything above is one payer reusing its own schedules across its own network products. It is not evidence that price position is a property of a practice in the market generally — that requires the same analysis on a different company's files, which we have not run. Read every number here as a statement about BlueCross BlueShield of Tennessee's July 2026 files and nothing wider.

What this changes about reading a benchmark

The practical version is short.

If you are checking your rates against "where the top of this payer's schedule sits," you are reading a label that moved by threefold between four of that payer's own products, reordered between services, and appears twice under different names. Which network's file the benchmark was built from is not a footnote. It is most of the answer.

The same is true from the other chair. A network-adequacy or rate-benchmarking exercise run inside a health plan that treats "the top of the schedule" as a single number is also reading four different objects, and the one it happens to have loaded decides the conclusion.

The position underneath is the more durable object — not perfectly durable, 0.60 is a long way from 1.0, but far more portable than the price attached to it. A practice that is expensive in one of this insurer's networks is usually expensive in the others, even when every specific number changes.

That gap between the two is the whole practical problem with public rate data right now. Anyone can download these files; increasingly, anyone can ask a chatbot for a price and get one back. The raw number stopped being scarce a while ago. What is still scarce is a neutral read of which file a number came from, whether the rate in front of you was negotiated or assigned, and where a specific practice actually stands once those two questions are answered. We do that assessment for either side of the table and as neither side's advocate — the benchmark, the position, and how both move when the next month's file publishes. What anyone does with the answer is theirs.

---

Method: BlueCross BlueShield of Tennessee in-network machine-readable files, _890 schema, vintage 2026-07-27, parsed from the raw published files. Fifteen home network products present; fourteen parsed in full and used (one, HMAS, carries a stale 2025 vintage with an empty first segment and was skipped). 2,462,194 provider-code rate rows persisted. Seventeen CPT codes across imaging, office visits, physical therapy, injection, and lab; the fifteen non-lab codes are used for all correlation figures, with the two lab codes reported separately as described. Global billing component only — billing_code_modifier is read, never dropped, so professional-only and technical-only rates are excluded rather than mixed. A provider's rate for a code is an observed rate present in the file, never an interpolated median between two values; an interpolating estimator would manufacture spurious differences between networks and inflate the discordant population. "Standard schedule" = a rate among the five most frequently repeated values for that network and code; "off-schedule" = not on those in either network compared; "discordant" = off-schedule in both and the two rates differ. Rate-identity percentages are computed over the same fifteen non-lab codes as the correlations. Rank agreement is Spearman correlation, computed per code and averaged across the fifteen; nulls are label shuffles on the identical population, three seeds. Overlap expectations under independence = N·(n₁/N)(n₂/N)(n₃/N); the schedule-stratified null permutes non-anchor codes within anchor-tier strata over 25 shuffles. All counts are provider (NPI) counts. No provider names, NPIs, or tax IDs are reported. Negotiated rates are contracted amounts and are not a guarantee of payment on any specific claim. Single payer, single state, single vintage — not a claim about any other insurer's files.