Pay per token. Nothing else.
SwyrmCompute is an OpenAI-compatible API served by gaming PCs. One price list for everyone: no subscription, no minimum monthly spend, no seat fees. You pay for the tokens our servers count, and nothing when a request fails.
Text models
| Model / API ID | Context | Input / 1M | Cached input / 1M | Output / 1M |
|---|---|---|---|---|
Qwen3.8-Flash-Nextswyrm/qwen3.8-flash-next | 262,144 tokens | $0.12 | $0.03 | $0.45 |
Qwen3.8-Flash-Next is a mixture-of-experts model served whole on each AMD GPU: about 50 output tokens per second, and follow-up messages reuse the conversation cache, so repeated context is billed at the cached rate and answers start in well under a second. Every node is benchmarked and certified for the context it can really serve (32K, 128K or 262K). Model details.
Images
Z-Image Turbo (6B parameters, Apache 2.0): $0.0025 per megapixel, billed in 512×512 steps rounded up, so a 1024×1024 image costs exactly $0.0025 and 1,000 of them cost $2.50. Sizes from 256 to 2048 px, PNG, JPEG or WebP. Failed, refused or filtered requests cost nothing. Images API.
What things cost
| Workload | Assumptions | Cost |
|---|---|---|
| Support chatbot | 1,000 replies, each with 2,000 tokens of prompt and 500 tokens of answer | $0.465 |
| Summarise a long report | a 100-page document (about 50,000 tokens) into a 1,000-token summary | $0.0065 |
| Coding session | 20 turns over a 30,000-token codebase, about 1,000 new tokens per question and 500 per answer; turns 2–20 reuse the conversation cache | $0.027 |
| Batch extraction | 10,000 product pages of 1,500 tokens each into 200-token JSON | $2.70 |
Computed from the price list above with Qwen3.8-Flash-Next. Token counts are approximate: about 750 English words make 1,000 tokens.
How billing works
- Each request reserves its maximum possible cost, then is settled on the tokens actually delivered; the difference goes straight back to your balance.
- If no GPU starts answering, nothing is charged. If a stream breaks midway, you pay only for what arrived.
- An
Idempotency-Keyheader makes retries safe: the same key never runs or charges twice. - Send
X-Session-Id(oruser) to keep a conversation on the same GPU and pay the cached rate for context it already holds. - Token counts come from our servers with the model's own tokenizer, never from the GPU that answered.
Paying
Your account has its own deposit address on Solana and on BNB Chain. Send USDT or USDC from any wallet or exchange (at least 1 USD per transfer) and the credit appears by itself, usually within a minute. The same balance pays for text and images. Holding SWYRM in a linked wallet for 7 days lowers the price by up to 6%; it is never required. Billing details.
Volume pricing
Running tens of millions of tokens a month or more, or thousands of images? Tell us your volume and we will come back with a price that fits. Talk to us
Frequently asked questions
How much does the SwyrmCompute API cost?
Qwen3.8-Flash-Next costs $0.12 per million input tokens, $0.03 per million cached input tokens and $0.45 per million output tokens, with up to 262K tokens of context. Z-Image Turbo images cost $0.0025 per megapixel. There is no subscription.
Is there a free tier or a minimum spend?
There is no subscription and no minimum monthly spend. A deposit needs at least 1 USD, and you only pay for delivered tokens.
Am I charged when a request fails?
No. If no GPU starts answering you pay nothing, and if a stream breaks midway you pay only for the tokens that arrived. Failed or refused image requests are free.
What are cached input tokens?
When a conversation continues on the same GPU, the context it already holds is billed at $0.03 per million tokens instead of $0.12. Send X-Session-Id (or the OpenAI user field) to keep a conversation on one GPU.
Can I pay by card?
Not for the API today. You fund your balance with USDT or USDC on Solana or BNB Chain, from any wallet or exchange.
Why is it this cheap?
The model runs on gaming PCs that already exist, owned by people who share their idle GPU. They earn 80% of what customers pay; there are no data-centre GPUs to pay off.