Whitepaper

Antimatter Compute, in plain words

Every model is a planet. Antimatter is the drive. This paper says what Antimatter Compute is, how each part of it works, and what it does not do. It describes only what is built and running at this address. Where something is not built, it says so. The field manual has the requests, the headers and the errors.

What Antimatter is

Antimatter Compute charts every language model it can reach, lets you talk to any of them, and then lets you use them through one OpenAI-compatible API behind a key your wallet mints. Antimatter is the short name, and the one the site uses.

Three surfaces share one catalog. The chart draws each model as a world you can fly to and read: its price, its context, what it has been measured to do. Chat puts any of them on the other end of a conversation, one at a time or several side by side. The API serves the same catalog to your code, with the same prices and the same measurements behind it.

The name is the idea. Every model is a planet, the wormhole is the way between them, and antimatter is the drive that makes the jump. The mark says the other half: antimatter is matter's mirror image, so the mark is one ring and its reflection across one line.

The problem

Language models come from many vendors. Each has its own key, its own bill, its own wire format and its own idea of an error. Prices move without notice. A catalog page says what a model supports, and the model does not always do it. Using several of them means writing the same integration several times and trusting claims nobody measured.

Antimatter answers with one door: one key, one base URL, one balance, and a chart of what is out there that shows what was measured and says so where nothing was. It asks for no card, no name and no email. A wallet signature is the identity.

The gateway

The gateway speaks two wires over the same catalog, keys and credits: the OpenAI chat completions wire at /v1/chat/completions and the Anthropic messages wire at /v1/messages. Embeddings are at /v1/embeddings. Both wires stream. Tool calling is translated in both directions, so a tool call written for one wire reaches a model that speaks the other. Images and structured output reach Anthropic's models too.

Routing

A model id names a world. An autopilot id names a policy, and the router picks the world:

  • antimatter/auto: price, measured latency and measured quality, scored together.
  • antimatter/auto:cheap: the lowest blended price.
  • antimatter/auto:fast: the lowest measured time to first token.

The score is computed per request over the worlds that fit. A request can steer it with the x-antimatter-route header: a policy, weights, and constraints such as a price ceiling, a minimum context, or a measured working tool call. An axis nobody measured counts for nothing, never for the average of the others, and the answer says unmeasured. POST /api/router/explain shows the decision for a request without sending it, and every answer names the world that served it in x-antimatter-model.

Fallbacks and breakers

Give a list of models and the gateway tries each in turn when one answers 429, fails with a 5xx or drops the connection. Before the first attempt it checks that every world in the chain has a published price, and it holds the worst case across the whole chain, because a fallback can cost more than what it replaces. Each upstream sits behind a circuit breaker: a run of failures opens it, traffic goes around it while it is open, and a trip on one instance is shared with the others.

Caching

Caching is off unless asked for. An exact repeat can be served from store. A thread whose earlier turns match byte for byte is recognised as the same thread, and a stored answer is only ever served for the whole conversation, never for a shorter one. A semantic layer can match a question that means the same thing, and a prompt too long for it to read in full is never a candidate. Anthropic's cache_control passes through untouched.

Your own key

Bring an OpenRouter, OpenAI or Anthropic key and the requests it serves bill your provider account, spend no credits, and come back with x-antimatter-byok: 1. A request on your key never falls back onto the house key: a plan that would is dropped and named in x-antimatter-route-warning. An answer your key paid for is never written to the shared cache. Provider keys are encrypted with a key held outside the database.

Clients

A TypeScript SDK, a Python SDK, a Go SDK, a command line client, an MCP server, and providers for the Vercel AI SDK and LangChain. The official OpenAI and Anthropic clients work as they are, pointed at this base URL, and a compatibility suite drives them against the gateway to keep that true. The whole surface is in the OpenAPI document, generated from the code.

Measurement

A chart is worth what its readings are worth. Antimatter measures, and where it has not measured, it says so.

  • Capability probes send a world small fixed tasks through the same path a customer's request takes, and record whether tool calling and JSON output actually work, what a long context does to recall, and how often an ordinary prompt is refused. Probes spend real money, so they run on a budget against a bounded set of worlds. A world never probed reads unmeasured, and a reading past its freshness window is marked as history.
  • A scheduled probe is the canary, and /api/health reports the newest one.
  • The router's quality axis is the probes plus the reliability measured from real traffic, each weighted by the evidence behind it. A world with neither reads unmeasured, and nothing is borrowed from another world to fill the gap.
  • Rankings count what this deployment was actually asked to serve. They measure adoption, not quality.
  • The flight recorder keeps one span per request: the world, the timing, the tokens, the cost and the outcome. Bodies are kept only for a key where you switch capture on, and only for the window you set. A captured request can be replayed against another world and the two answers read side by side. Spans export over OpenTelemetry, and alert rules watch them.

Why nothing on the site is invented

A deployment with no upstream key shows an empty catalog and answers 503, never a mocked reply. A reading nobody took reads unmeasured, never zero. A readout whose source has not answered is not shown at all, rather than shown with a placeholder. Prices come from the upstream catalog and from the vendors' published tables.

Identity and keys

You sign in by signing a message with a wallet. The message says what it is, the signature proves you control the address, and it authorizes no transaction and costs nothing. The nonce it signs is bound to that address and used once. The session is a signed cookie on this site. There is no password.

Keys are minted from a signed in session, on the dashboard or the bridge. A key starts am_live_, is shown once, and is stored only as a SHA-256 digest, so nobody here can read it back. A key carries scopes, an optional expiry and an optional IP allowlist, and a policy can hold it to certain models and to a daily or monthly limit. Keys rotate in place, and an account can cap its own monthly spend.

Organizations hold one shared balance. A lead sets a budget for each member, roles are enforced on the server, and a key minted for an organization belongs to it rather than to whoever minted it. Every action that touches money or security is written to an append only audit log.

Money

Antimatter is funded with stablecoins only. There are no card payments. You send USDC or USDT to the deposit address the deployment publishes, on an accepted network: Base and Arbitrum One by default, and Solana where the deployment names a Solana address. Then you claim the transfer by its transaction hash, or on Solana by its signature with a short message signed by the sending wallet. A transfer credits exactly once, however many times it is claimed. A stablecoin's decimals come from configuration and are never guessed. A deployment with no deposit address takes no deposits, and says so.

The ledger is kept in whole micro USD. Before a request goes out, the worst case it could cost is held against the balance. When it settles, the hold becomes the actual cost and the rest is released. A stream the client cuts short is billed for what was read, the prompt included. Receipts are read from the ledger, and monthly statements are built from it, on the billing page.

An account that has never deposited gets a lifetime free allowance, set by the operator. Past it, the API answers 402 until the account is funded. A house breaker caps what free use can spend in a day, across everyone. Credit cannot be withdrawn: it is spent on requests.

The chart and the ship

Every model in the catalog is drawn as a planet. Its look is derived from its id, so a world keeps the same face on every visit and every screen, and a vendor's worlds gather into one system.

The chart is flown, not scrolled. The ship is a starship with warp nacelles, and it moves like one: it keeps its velocity until the engine changes it, turns round to brake, and arrives standing still. Hold a world under the scan and it joins your codex. Voyages are routes through the chart with a line about each stop, shared as a link. Seasons are windows with a goal, and a standing that cannot change once the window closes. Photo mode takes a picture of where you are. Other explorers show on the chart while they fly. Charting earns hulls and trails, and they are cosmetic: nothing you earn changes a price or a limit.

Each world has its own page with its prices, its context, its measured latency, its capability readings and a live line to it. Each vendor has a system page. The observatory draws your own recent use as a solar system.

Privacy

What is stored

  • Your wallet address, and your sessions.
  • Your keys, as digests only. Provider keys, encrypted with a key held outside the database.
  • One usage row per request: the key, the world, the upstream, the status, the tokens, the cost, the timing, and any session label you sent. Not the prompt and not the reply.
  • What you make here: presets, aliases, saved prompts, webhooks, alert rules, organizations, and your chart of the galaxy, with a public profile only if you publish one.
  • Chat threads: in your browser while you are signed out, and on the server in plaintext once you sign in, so they follow your wallet.
  • Request and response bodies, only for a key where you switch capture on, and only for the window you set.
  • Your position on the chart while you fly it, briefly, so other explorers see a light.
  • The IP address a security or money action came from, in the audit log.

Export and erasure

Export gives you everything held against your wallet as one JSON file, from the dashboard. Secrets are the only omission, because they cannot be read back.

Erasure deletes the account, through DELETE /api/account with a confirmation phrase. Two things outlast it on purpose. Ledger rows lose their link to you but stay, because the money has to add up and a deposit must never be claimable twice. And the free allowance already spent is remembered under a one way key of your address, so deleting an account does not buy a second one. Erasing an account that still holds credit forfeits that credit: the answer names the amount, and nothing is erased until you confirm the forfeit.

The token

$ANTI is the Antimatter token. It launches on Sable, a launchpad on Robinhood Chain. It has not launched. There is no contract address yet, and until one is published here, no address is it.

What it does

One thing, and only once the operator switches it on. An account whose own wallet holds at least a set amount of $ANTI on Robinhood Chain gets a larger lifetime free allowance. The wallet you sign in with is the one that is read; there is nothing to link. The amount and the size of the bonus are the operator's settings. Nothing else changes: prices, limits and routing are the same for everyone.

What it is not

It is not a way to pay. Antimatter is funded with stablecoins only, and $ANTI buys nothing here. It is not a share of anything: it carries no claim on revenue, on ownership or on a vote. There is no promise about its value. This site states no price, holder count or market cap that a source did not measure.

A token is a risk you take alone. A token can lose all of its value, and Sable's own page says so. Read it before you do anything.

What is not built

Said plainly, so nobody has to find it out.

  • The token has not launched, and the holder perk stays off until it has and the operator switches it on.
  • There are no card payments and no fiat. Funding is stablecoins, claimed by transaction.
  • Credit cannot be withdrawn to a wallet. It is spent on requests.
  • Antimatter hosts no models. Every answer comes from an upstream provider, and a world is only as available as the provider behind it.
  • There is no benchmark of its own and no model judging another. Quality is what the probes and real traffic measured, and nothing more.
  • The API does not remember conversations. Each request carries its own messages; only chat, on the site, keeps threads.

This page describes what runs today. It makes no promise about what comes next.

Antimatter