fusegrid

A spend ceiling that holds under concurrency. v0.3.2 · API docs · health

The result, in one request

curl "{HOST}/demo/overrun?concurrency=20"

Runs the same 20-request burst twice against a $0.10 ceiling that permits 5 calls: once with the budget checked after the call, once with a reservation taken before it. Post-hoc accounting overruns. Reserve-then-settle refuses the 6th request.

Or spend a budget down yourself

curl -X POST {HOST}/v1/chat/completions \
  -H 'content-type: application/json' \
  -d '{"model":"demo-expensive-1","budget_key":"demo-tight",
       "max_tokens":256,"messages":[{"role":"user","content":"hello"}]}'

curl {HOST}/v1/budgets/demo-tight     # watch remaining_usd fall
curl -X POST {HOST}/demo/reset        # put it back for the next visitor

Keep going and the ceiling refuses you with 429 and a budget_exceeded body carrying remaining_usd and ceiling_usd — before the spend, not after it.

Budgets on this instance

budget_keyceiling
demo$0.50
demo-tight$0.05
demo-generous$5.00

Models

modelinput /Mtokoutput /Mtok
demo-expensive-1$20$40
demo-cheap-1$1$2

Priced far above real models on purpose, so a ceiling is reachable in a handful of requests rather than a few thousand. An unpriced model is refused, never charged at zero.

The upstream here is a stub, and that is deliberate

fusegrid does not claim to forward requests well — anything does that. It claims a ceiling holds. Demonstrating that needs a ledger, a pricing table and concurrency; it does not need a real model, and wiring one in would mean publishing a credential or asking for yours. The reserve → call → settle path runs in full and every number above is the real ledger's.