Skip to content
Checking statusStatus

Balance & billing

Credit packs, per-token rates, prompt caching and what a request costs.

How billing works

  • Access is prepaid. Pick a plan and buy a credit pack in the dashboard:
    • Claude (every Claude model, including Fable 5.2 and Opus 5.5), ChatGPT (every ChatGPT model, including GPT-6 Astra) or Gemini (every Gemini model, including Argon 4): $50 for $20, $100 for $30, $200 for $40, $400 for $55, $1,000 for $95, $2,000 for $150.
    • Two families on one key, Claude & ChatGPT, Claude & Gemini or ChatGPT & Gemini: $50 for $25, $100 for $35, $200 for $45, $400 for $65, $1,000 for $100, $2,000 for $155.
    • All three, Claude, ChatGPT & Gemini: .
  • Once your payment confirms, you get an email, and your API key follows by email shortly after. Your order's status is always visible in your dashboard.
  • A key only calls the model families in its plan. A request for a model outside the plan is rejected before it is routed.
  • Your key's credit is checked before each request runs. When it is used up, requests fail with insufficient_balance.
  • Each request's token usage is charged to your key's credit at the rate of the model that served it. See How routing works.
  • To top up, buy another pack from your dashboard and we add the credit to your key.
  • Payment methods: USDC & USDT, Bitcoin, ETH and 100+ coins.

Rates

USD per 1M tokens. Most models in a tier bill identically; the newest (Claude Fable 5.2, Mythos 5, Opus 5.5, Haiku 5.5 and GPT-6 Astra) bill at 1.5x the tier rates below, and each model page shows its exact rate. See which models are in each tier on the Models page.

TierInput5m cache write1h cache writeCached inputOutput
Frontier$3.60$3.60$3.60$2.10$24.00
Balanced$2.00$2.00$2.00$1.20$10.00
Speed$0.80$0.80$0.80$0.50$4.00

Prompt caching

When a request reuses a prompt prefix the model has cached, those input tokens are billed at the cached-input rate instead of the input rate. Writing to the cache is billed at the input rate.

  • Claude models cache explicitly: mark breakpoints with cache_control on the Messages API, with a 5-minute or 1-hour lifetime.
  • GPT models cache long, repeated prompt prefixes automatically.

Worked example

A Frontier-tier request with 20,000 input tokens, of which 15,000 come from the cache, and 2,000 output tokens:

LineTokensRate / 1MCost
Uncached input5,000$3.60$0.02
Cached input15,000$2.10$0.03
Output2,000$24.00$0.05
Total$0.10

Without caching the same request would cost $0.12. Try your own numbers in the cost estimator.