Loot Get a key

Cheaper than
the vendor

A metered, OpenAI-compatible API for chat, embeddings, images and video, resold below what OpenAI, Google and Anthropic charge you directly. Prepaid in USDC or $ROB on Solana. No account, no card, no KYC — the first request is closer than the signup form.

Live now: chat, embeddings, images, video. Try any model in the browser.

A man in a suit at a vending machine, feeding it money.
Text, vs official−5%
Media, vs officialList
Models priced—
Deposit fee0%

Live from the ledger

Counted from the ledger's own rows, not a counter somebody increments, and refreshed once a minute. Usage and adoption only: what we earn is not on this page.

Requests served—
Tokens processed—
Images generated—
Videos generated—
$ROB accepted—
Keys minted—
Wallets signed in—
Models live—

Check it in thirty seconds

Per million tokens, input and output. The right-hand column is the closest competitor, which prices at official plus five. We price at official minus five, so every shared model is 9.5% under them and under the vendor at the same time.

Model Official Loot vs official minirouter vs minirouter
——————

Rounded to the cent for reading. The pricing page carries the full figure, and billing carries six decimals — nothing here is rounded before it reaches your balance.

The discount is a dashboard rate, not a contract. Our upstream sells at 20% below list, and every price above is that cost plus our margin — not a hardcoded number. If the rate moves, the table moves with it, because prices are stored per model and re-harvested rather than typed in. Only one of the two competitor rows above is a published price; the rest are their +5% rule applied, and worth re-checking before you rely on them.

DeepSeek, after two in the afternoon

DeepSeek V4 Pro is half price upstream outside 00:00–14:00 UTC, and that discount is passed through rather than pocketed. Our margin stays where it always is; your bill halves. The API tells you which window you are in and what the same call costs in the other one.

Official
—

input, per 1M · — output

Loot, 00:00–14:00 UTC
—

input, per 1M · — output

Loot, the other ten hours
—

input, per 1M · — output — roughly half of official

The half a text router does not have

Images, video, audio and embeddings on the same key, the same balance and the same base URL. Media sells at list rather than under it — there is no competitor undercutting it, and pretending otherwise would be a discount taken out of the margin that pays for failed generations.

Chat

Streaming SSE, cost in the final usage chunk. Live.

Embeddings

Synchronous, same path, same billing. Live.

Images & video

Submit, follow the task, and it settles. A failed generation is not billed. Live, and every model has a playground.

Audio

The route exists, but no speech model publishes a price upstream, so none is sold yet.

Point a client at it

It is the OpenAI API. Every client that speaks it already speaks this — and both base-URL forms work, so there is nothing to get wrong.

# both of these are the same endpoint. no bold warning required. curl …/v1/chat/completions \ -H "Authorization: Bearer loot-live-…" \ -H "Content-Type: application/json" \ -d '{"model":"gpt-5","messages":[{"role":"user","content":"hi"}]}' < x-loot-cost-usd: 0.00002100 < x-loot-balance-usd: 4.97891200 < x-loot-ttft-ms: 412

What you should know before you send money

One upstream. Every model here is resold from a single provider. If it is down, Loot is down — there is no second route to fail over to. That is an accepted launch trade, stated rather than discovered.

No withdrawals. A balance funds calls and there is no path back out of it. Send what you intend to spend.

We host what you generate, and we moderate it. Generated images and video sit on our domain, which is an abuse-liability problem text routing does not have. Every generation is classified before it is stored, every file carries a takedown address in its own headers, and nothing is kept past the retention window.

Storage has a retention window. Generated media is deleted on a stated schedule, published before the first byte is stored rather than after the bill arrives.