Overview
Where the money went
Every request through Fabric, priced against what it would have cost with the levers off. The difference, attributed lever by lever, is the product.
Requests
—
Baseline
—
with every lever off
Actual
—
what we paid
Saved
—
Cache hits
—
Spend waterfall
pipeline order · each lever credited what it personally causedSavings are order-dependent: a lever running second is credited only with the delta it caused given what ran before it. The segments sum exactly to baseline − actual.
Compressors
shares of the compression lever · not sumsThese counts are shares, not sums — never add them up. Headroom measures a compressor on only the segment of the prompt it was handed, which is smaller than the whole request and different from compressor to compressor. The rows below are therefore a relative breakdown of the compression lever total (tokens_before − tokens_after); they do not sum to the request delta.
Loading…
Prefix cache health
provider-side cacheLoading…
On real traffic the provider's prefix cache is the largest single line item, far larger than compression. A busted prefix means the provider re-reads and re-bills the whole prefix, which is why this page gives it space proportional to its cost.
Realized by lever
No realized savings yet.
Input tokens saved—
Output tokens—
By workflow
No requests carry an X-Fabric-Workflow label yet.
Recent requests
| when | workflow | served | baseline | actual | saved | latency |
|---|---|---|---|---|---|---|
| No requests in this window. | ||||||