Briefing

Every model is a world. Learn to choose one.

What the worlds are, what they cost, what we charge, how they differ and how to pay, in one short flight. Scroll to fly.

01Worlds

Every model is a world

This is Mistral: Mistral Medium 3, a world from the middle of the price range. Every language model charted here is drawn as a world, its face grown from its id, so it looks the same each time you come back.

466 worlds from 57 makers are charted here. The largest systems are openai with 107, qwen with 53 and google with 44.

One key reaches all of them. One base URL, one balance, one bill, through an API that speaks the OpenAI and Anthropic wires, so the clients you already use work as they are.

02Cost

What a world costs

A world is priced per million tokens, and what you send in and what comes back out are priced apart. A token is roughly three quarters of a word, so a million tokens is about 750,000 words.

Above the line, what 427 priced worlds charge for input. Below it, in the mirror, what they charge for output. Output usually costs more: on the typical world here a token out costs 4.1 times a token in. 20 more are published at no charge, and are throttled upstream.

The range is wide. The cheapest, Mistral: Mistral Nemo, asks $0.021 in and $0.033 out. The one in the middle, Mistral: Mistral Medium 3, asks $0.44 and $2.20. The dearest, OpenAI: o1-pro, asks $165 and $660.

So the same work costs very different amounts. Computed from those prices:

What two example tasks cost on three worlds
TaskCheapestMiddleDearest
A chat turn1,000 in, 500 out$0.000037$0.0015$0.49
A long report, summarised30,000 in, 1,000 out$0.00066$0.015$5.61

03Our margin

What we charge

Every price here is the upstream's own price plus a margin of 10 percent. That margin is how Antimatter is paid, and it is kept small on purpose.

Mistral: Mistral Medium 3 costs $0.40 per million tokens in at its source, and $0.44 here. Out, $2 at the source and $2.20 here.

The price you see is the price you pay. The chart, every world's page, the API's model list and your bill all read the same number, margin included. Nothing is added at the end.

04Differences

How worlds differ

Context is how much a world can read at once, your words and its answer together. The smallest window here holds 4,095 tokens, about 3,100 words. The one in the middle holds 262,144. The largest, SpaceXAI: Grok 4.20 Multi-Agent, holds 2,000,000, about 1,500,000 words. 179 worlds read a million tokens or more.

Speed is measured here, not quoted. Time to first token is the wait before an answer starts. Over the last seven days Mistral: Mistral Small 3 started in 247 ms at the median, and Nex AGI: Nex-N2.5-Mini in 2.3 s. Each comet crosses in a time proportional to its world's wait.

Abilities differ too. 295 of 466 worlds take images as input. Tool calls were measured working on 5 of the 7 worlds tested, and structured output on 7 of 7. Every world's page shows its readings, and says unmeasured where nobody measured.

05Autopilot

Let the autopilot choose

Not sure which world? Ask for antimatter/auto instead of a model and the router picks one for each request on price, measured latency and measured quality, scored together.

Or say what matters. antimatter/auto:cheap takes the lowest blended price, and antimatter/auto:fast the lowest measured time to first token. Task modes add hard requirements, read only off evidence:

  • :code code
  • :long long context
  • :vision vision
  • :tools tool calling
  • :json structured output

A failed world need not fail the request. Send x-antimatter-fallback: comparable and a failure is handed to the nearest world that is no dearer on either price, chosen before the request goes out so your hold covers it. Every answer names the world that served it.

06Paying

Paying

Start free. A new account gets $0.25 of use on the house, once, before it deposits anything. Sign in with a wallet: no card, no name, no email.

Fund your balance with stablecoins: USDG, USDC or USDT on Robinhood Chain, Base or Arbitrum One. Send, then claim the transfer by its transaction; a transfer credits exactly once. There are no cards.

You pay for what you use. While a request runs, its worst case is held against your balance. When it settles you are charged what it used, to the millionth of a dollar, and the rest comes back. Credit is spent on requests and cannot be withdrawn.

Antimatter