[ dispatch ]

Dispatch, our real-time routing engine. Every provider competes, the cheapest wins.

Every request goes to the cheapest provider for that model. You pay below the source price. No platform fees, no top-up fees.

// what it does

Every request is a live competition.

01 / Quote

Every provider quotes

Each provider that serves the model you ask for quotes on your request. A quote that breaks one of your rules drops out.

02 / Win

The best quote serves it

The best quote for your mode wins. If that provider fails, times out or turns the request down, the next quote takes it before any token reaches you.

03 / Earn

The margin is network revenue

The margin on every request is network revenue. Half goes to USDF holders, every hour.

// modes

Four modes. Auto picks for you.

Coming

cheapest (default)

Cheapest

The lowest price among providers that are working now.

fastest

Fastest

The highest measured output speed. Within 1 token a second, the cheaper one wins.

smartest

Smartest

The route most faithful to the model: its developer's own API first, then full-precision weights, then the most reliable.

auto

Auto

Auto picks the best model for every prompt, at the lowest price.

Send mode in the request, or add it to the model name after a colon: deepseek-v3.2:fastest. The word counts as a mode only when it is the whole text after the last colon.

// rules

Rules on every request.

Coming

Add rules to any request. A provider that breaks a rule does not compete.

RuleWhat it does
price_capThe most the whole request may cost, in USD. If it cannot be met, the request is refused before anything is charged.
max_ttft_msThe longest wait for the first token. A stream that has not started in time moves to the next provider.
min_tpsThe slowest output speed you accept, in tokens a second. Slower providers go to the back.
regionOnly providers that process in this region (us, eu).
zero_retentiontrue: only providers that keep nothing of your request.
no_trainingtrue: only providers that never train on requests.
providersAllow and deny lists of providers.
modelsAllow and deny lists of models, with * for any text. For auto, the models it may choose from.
cost_tierFor auto: economy, balanced or premium.

Coming The price is locked. Before the first token, the request's worst case is held: that is its price cap, and you never pay more. You pay only for the tokens actually used.

// receipt

A receipt on every answer.

Coming

Every answer carries a receipt in receipt.router, with the same facts in its headers.

FieldWhat it says
modelThe model that answered.
price_capThe most the request could have cost. Held before the first token.
chargeWhat it cost: only the tokens actually used.
savingWhat you saved against the source price, in USD and percent.
providers_competedHow many providers quoted on the request.
providerFoundry network.

// privacy

Zero retention by default.

Nothing kept

By default a request goes only to providers whose published policy keeps no request data. Only routes that promise it can win.

Your rules, on top

Add no_training to rule out any provider that trains on requests, or region to keep processing in the US or the EU.

How privacy works

// why dispatch

Why it beats buying direct and other routers.

Buying directOther routersDispatch
What you payOne provider's list priceThe source price, plus a fee to add fundsBelow the source price. No platform fees, no top-up fees.
Who serves itOne provider, whatever its price todayA provider order you set or a default listEvery provider that serves the model quotes, and the best quote wins
If a provider failsThe error is yours to handleFallbacks, often set up by youThe next quote takes it before any token reaches you
Your dataWhatever that provider keepsZero retention as an opt-inZero retention by default
What it can costKnown after the billA ceiling you can setThe worst case is held before the first token. You never pay more.
ProofA monthly invoiceA usage logA receipt on every answer
Your balanceCredit with one providerCredit on one platformUSDF: redeemable 1:1, never expires, earns every hour

The first two columns describe the usual terms of each way to buy, not any one product.

Compare Foundry with another router, row by row, or read the Benchmarks.

// example

Try it with one request.

Coming

Auto picks the model, the request never costs more than 2 cents, and only zero-retention providers can serve it.

Auto with rules

curl https://api.usdf.fi/v1/chat/completions \
  -H "Authorization: Bearer $USDF_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "auto",
    "messages": [{ "role": "user", "content": "Summarise this contract." }],
    "rules": { "price_cap": 0.02, "zero_retention": true, "cost_tier": "balanced" }
  }'

Every model.
The best quote wins.