Token-based pricing, at our cost
Prices are our wholesale cost plus 5% to cover payment fees. Per 1M tokens, live from our billing catalog. Pay with any credit card via Stripe.
Model pricing
Per 1M tokens, cheapest first — live from our billing catalog
| Model | Input / 1M | Output / 1M |
|---|---|---|
Models from each provider's official releases, with the provider's own model IDs. See the model catalog for the full list.
Image · Audio · Video · Embeddings
Billed at our cost price. Video is async (create → poll → download); polling and downloading are free.
Video is billed per generated second; seconds is 4 / 8 / 12. Full model list & examples in the docs.
No per-seat. No minimums. No markup.
Pay for forwarded traffic at our cost. Balance never expires.
FAQ
Are these the real, production Chinese models?+
Yes — DeepSeek V4, Qwen3, GLM-5, Kimi K3, Doubao Seed 2.0 and more are each provider's official release, with the provider's own model IDs.
How reliable is this?+
Most models are served over more than one service line: if a line is busy or down, the request automatically fails over to a backup line — no retry needed on your side. Tenants can also opt in to equivalent-model substitution for when a model itself is unavailable.
Do I need a Chinese entity, phone number, or visa?+
No. Sign up with any email. We handle provider-side compliance. Pay via Stripe in USD.
Can I still use GPT-4o / Claude / Gemini if I need them?+
Yes — same API. All models billed at our cost price (wholesale + 5%). Use the `model` parameter to pick, or set `model="auto"` and let our router choose the cheapest.
What does "cached" pricing mean?+
Some models charge less for prompt tokens they've already processed (e.g. a system prompt or long context you resend every turn). When a model supports it, the cached rate is shown next to its input price and applied automatically — no extra parameter needed.
Why doesn't my invoiced unit price exactly match the list price?+
Two reasons: some models price by input length in tiers, and the price shown here is the lowest tier — longer inputs bill at a higher tier. Also, a model may have multiple service lines, and each request is billed at the price of the line that actually served it, so the invoiced rate can differ slightly from the list price.
Data retention?+
Request bodies are forwarded, not persisted. Only anonymized metadata (token counts, model used, latency) is stored for billing and analytics.
How do I migrate from OpenAI?+
Change `base_url` in your OpenAI SDK to `https://promptoll.com/v1`. That's it. Your existing code keeps working. Switch `model` to a cheaper one when you're ready.