Dolphin
tier/dolphin
Open-weight models served at no charge. Rate-limited, never metered.
PAYMENT · NONE
SETTLES · —
A router, not a reseller
We put one OpenAI-compatible endpoint in front of every major model and charge for it the way the web was originally supposed to charge for things — per request, at the moment of the request.
Powered by x402400+ modelsSettled in USDG on Robinhood Chain
We did not set outto build a gateway.
We were building something else, and we kept writing the same four hundred lines of glue: a key vault, a retry policy with exponential backoff, a per-team quota table, a cron job that emailed somebody when a prepaid balance dropped below a threshold. None of it was product. All of it was necessary, because every model API assumes an account, and an account is a small piece of infrastructure you are now responsible for.
Then one of our upstream providers had a bad afternoon and our product had a bad afternoon, and the postmortem action item was the one everybody writes: add a second provider. Which meant a second account, a second key, a second rate-limit tier and a second rate card. The glue doubled.
Wuraw is what happens when you decide that glue is the product. One endpoint. One request shape. The model-by-provider matrix lives behind it, the failover lives behind it, and the payment lives inside the request instead of in a billing relationship that had to exist before you sent it.
The unusual part is not the routing — routers are well understood. It is that the thing we removed was the account. Not softened, not streamlined: removed. There is nothing to sign up for because there is nothing for a sign-up to create.
41 hours a year,
or about 12 minutes.
LLM providers run at 99–99.5% uptime — 6 to 14× worse than cloud infrastructure.
That gap is the entire case for a router, and it is arithmetic rather than opinion. Two providers with automatic failover cut expected downtime from about 41 hours a year to about 12 minutes. Forty-one hours is two and a half thousand minutes of your product being down for a reason your users will attribute to you and not to your vendor.
Everything else a gateway sells you — the unified catalog, the single rate card, the cost-optimized routing policy — is convenience. This one is availability, and it is worth more than all of the convenience combined. It is also the reason we run a redundant tier over the same catalog rather than a single cheapest path: the second provider is not there to be cheaper, it is there to be up.
Source: TianPan.co, LLM API Resilience in Production · 2026-03-11
~41h
Expected downtime per year at 99.53% availability, no failover path.
~12m
Two independent providers at the same availability, automatic fallback between them.
Three tiers
one catalog.
A request names a model. The router decides which upstream serves it, using the policy you set and the health it observes.
Free
Surplus
tier/surplus
Frontier models at list price. Each call settles in USDG before the stream opens.
PAYMENT · X402
SETTLES · USDG / ROBINHOOD
AntSeed
tier/antseed
Redundant capacity for the same catalog — the fallback when an upstream degrades.
PAYMENT · X402
SETTLES · USDG / ROBINHOOD
The free tier is not a trial. It is rate-limited open-weight capacity that never touches the chain, which means it also keeps working during a settlement outage — the one failure mode a payment-native router has that a credit-based one does not.
Four rules we
have not broken yet.
01
No account, therefore no gate
Every geographic block, verification step and suspension policy in this market attaches to an account. Remove the account and the attachment point goes with it. A signature over an amount is the whole relationship.
02
Your money stays yours until the call
We hold no float. There is no balance to strand, nothing to reclaim when you leave, and no custodial failure mode. Payment authorizes exactly one transfer of exactly one amount, at the moment the request is made.
03
Cite it or do not claim it
Every number on this site carries a source and a date. Where the sources disagree — and on this subject they often do — we publish the qualifier next to the figure instead of picking the flattering read.
04
Vendors change, your endpoint does not
Aliases sit between your code and a provider SKU. When a model is retired, the change is a reviewable config diff against a name you own, not a search for every call site in the repository.
A short timeline
of removing things.
2025
The 402 thesis
Coinbase revives HTTP 402 as an open payment standard. We are running a small internal model proxy at the time and realize the thing we keep rebuilding — key custody, per-team quotas, top-up alerts — is all compensation for a missing payment primitive.
2026 · Q1
One endpoint, three tiers
First working router: a single OpenAI-compatible surface in front of a free open-weight tier and two metered upstreams, with fallback between them. No payment yet — just the routing table and the failover logic.
2026 · Q2
Payment moves into the request
Paid routes start answering 402 with machine-readable terms instead of 401 with a key hint. Settlement runs through a facilitator against a gasless EIP-3009 authorization. The account model comes out entirely.
2026 · Jul
Robinhood Chain mainnet
The chain goes live on 2026-07-01 with USDG as its first natively issued stablecoin. Gas is paid in ETH, so a payer funds one asset and watches one balance. We make it the default settlement network.
2026 · Now
Catalog and budgets
400+ models across 3 upstream tiers, 19 settlement networks accepted, and per-agent spend caps enforced at the signer rather than in an admin console.
The x402 InnovationInitiative.
The x402 Foundation went operationally live under the Linux Foundation with 40 member organizations. Members include Visa, Mastercard, American Express, Stripe, Google, AWS, Shopify, Cloudflare and Coinbase. A payment standard with that membership list is no longer a crypto experiment; it is an interface that the people who operate the card rails have agreed to look at.
We run an open initiative for teams building on the same rail. It is not a grant programme and there is no token allocation attached. What we offer is concrete: a settlement integration reviewed by people who have already shipped one, early access to catalog and routing changes before they land, and a named channel for the awkward parts — facilitator failover, header-version drift, refund semantics when an upstream returns a 5xx after payment has settled.
Source: Linux Foundation / PRNewswire · 2026-07-14
- Agent frameworks that want per-caller spend caps enforced at the signer.
- Facilitators and wallet teams working on EIP-3009 authorization flows.
- Inference providers with spare capacity who want demand without a billing relationship.
- Anyone shipping a paid HTTP resource who has hit the same header-versioning wall we did.
Applications open through the contact channel listed in the footer. We answer every one, including the declines.
Keep reading.
Documentation
Quickstart, the endpoint contract, routing policies, error codes and rate-limit tiers.
Open
Whitepaper
The architecture, the settlement model, the security assumptions and what we cannot promise.
Open
Blog
Longer arguments — 402 history, gateway benchmarking methodology, model-retirement drills.
Open