Recent cases first. Pick one, or search by case id, title, or reference number.
| Title | Reference | Stage | Bidders | Started |
|---|---|---|---|---|
| Loading… | ||||
One row per (service, model) in this case's LLM call log — click the service to open it bidder-wise, and a bidder to open its sub-categories. Tokens are the recorded counts; thinking is billed at the output rate; a row with no usage recorded is estimated from bytes and says so. Rates are the form's.
| Service | Model | Calls | Attempts | Prompt tok | of which cached | Answer tok | Thinking tok | Basis | Input $ | Output $ | Line $ | Console line |
|---|
| Service | Model | Calls | Failed calls | Re-run calls | Attempts | Billed | Not billed | Attempts by outcome | Retries $ | Cache saved $ |
|---|
Modelled from this case's recorded quantities: GCP bills Cloud Run by instance-second, never by case, so each line shows the assumption it rests on. The total replaces the flat allowance in the Measured column and prefills the predicted input.
| Item | Basis | $ |
|---|
| Line | Predicted | Measured |
|---|
| Flow step | Doc AI | 3.5-flash | 3.6-flash | flash-lite | Embed + BQ | Step total | Share |
|---|
Where the money goes — this run's measured quantities bucketed by the step the user experiences, priced at the current rates. — = the step never ran.
What a month costs at a given case load: processing per case (LLM, Document AI, embeddings, BigQuery) and platform per case (Cloud Run time, storage, tasks) scale with N; the always-on floor does not. Per-case figures are the calculator's current inputs — open a representative case first.
| Item | Basis | Full load | Half load |
|---|
What this environment costs per month when nothing runs and everything has quiesced: warm minimum instances, the database, the load balancer, storage. Cloud Run services at min 0 cost nothing idle.
| Item | Basis | Monthly $ |
|---|
What this environment would cost per month with every Cloud Run service at its maximum instances, CPU and memory for the whole month. The database stays at its tier (that is its ceiling); requests, load-balancer data and egress are assumptions you can edit.
| Item | Basis | Monthly $ |
|---|
Ceilings and floors at list price: free tiers and committed-use discounts are ignored. The Cloud SQL instance, load balancer, VPC/DNS, registry and the pricing console are project singletons shared by dev, UAT and prod; they are listed on every sheet with an env + ⅓ share total.
Completed cases started at or after the cut-off. Spend on cases that never completed shows separately — it was still paid for.
| Flow step | Doc AI | 3.5-flash | 3.6-flash | flash-lite | Embed + BQ | Step total | Share |
|---|
Fleet totals use the same client-side rates as the calculator. Wasted spend is priced per model from recorded tokens (costing record v4); an older record has only logged bytes, and its wasted spend is then a floor and says so. Response is cached server-side for 10 minutes.