One connector finished ~2.5× more of what was asked.

Figure 2

What each question cost, both sides

One shared scale, one row per question. The vertical rule is the floor: the publisher MCP never answered anything for less than 143k tokens, because a 100+-tool schema goes into every call whether the question needs it or not. Skai's cheapest question cost 6k.

Each bar is the median of 10 runs of that question, on both sides, in one round of testing. The vertical rule is the cheapest question the publisher MCP answered all round — nothing it did cost less than that, because a 100+-tool schema is carried into every call whether the question needs it or not. Skai's side has no such floor: its cheapest question in the battery cost 6k. The 4 multi-turn flows in the same battery are held out of this chart — they chain several questions into one session, so they aren't comparable to a single ask.

Figure 3

The same cost, projected past what we measured

Skai's per-query cost barely moves as accounts are added. Publisher MCPs carry their full schema on every call, so the gap widens with scale rather than narrowing. Measured head-to-head cost per query runs 1.5× against small read-only publisher schemas and up to 8× against a large one, corroborated in this round at 4.1×–5× the tokens per query.

Skai's scaling exponent — measured

The same cross-account question asked against 1 account and against 10 under one login: 58,607 vs. 74,602 tokens per query. A 27% rise for ten times the accounts, because a cross-account question costs about one extra call however many accounts are in scope.

~0.10

The publisher exponent — modeled

An engineering assumption, not a regression. We ran the identical 1-versus-10 test on that side and could not use the result: a prompt-wording anomaly on the single-account run, plus up to 48% day-to-day variance on repeated identical queries. It stays as the prior modeled figure rather than being replaced by an unreliable measurement.

0.6

Where the curve is anchored

Both sides start from directly observed per-query averages at 2 accounts. Everything to the right of that is the model, not a measurement.

2 accounts

Plotted at a reference footprint of 3 publishers × 30 accounts and 300 reporting questions a month — 90 accounts in scope, marking $39.62 against $394.12 a month. Run it against your own footprint on the other tab.

How it was run

Two rounds. Round one: five reporting workflows, run twice each — once through Skai's Reporting MCP, once calling a publisher's MCP directly — against publishers whose own MCPs ship small, read-only tool schemas. Round two: 27 real reporting and diagnostic tasks, 10 runs each, in two account-scoping configurations of a publisher MCP shipping 100+ tools across reporting, account management, and campaign management, graded independently by three judges — one automated, two manual — over roughly 2,300 ballots.

“Real, data-backed answer” means judged completeness above the rubric floor: not a bare refusal, not an unresolved clarifying question standing in for the whole answer, not an empty or errored response. Fig. 2's bars are the median of the 10 runs of each question on each side, from the publisher MCP's stronger account-scoping mode rather than its weaker one. Pricing throughout is Claude Sonnet 5 at $2.00/MTok in and $10.00/MTok out, cache tokens excluded.

Back to the interactive view to run the cost model against your own footprint.