Blog ·
How the price stays low
After the beta we charge $0.005 per million input tokens, $0.001 per million cached input tokens and $0.14 per million output tokens. Here is where that comes from, and what you give up for it.
Older hardware
We run the model on older GPUs. Hosts that sell speed buy the newest cards and price them in. Older cards cost much less, and we pass the difference on.
One model
We serve DeepSeek V4.1 Flash 1017 and nothing else. Every card works on the same model, so none sits idle waiting for traffic on another one.
Waiting instead of refusing
There is no rate limit on your key. When we're busy, your request waits and is retried for up to 30 seconds instead of failing with a 429. Keeping the cards full most of the time is what makes a low price possible.
The prompt cache
When a request starts with the same text as an earlier one, that part comes from cache, and cached input costs $0.001 per million tokens. Coding agents resend the whole conversation on every turn, so most of their input is a cache hit (90% in the agent day we price in the comparison).
What you trade
- Under heavy load a request can wait up to 30 seconds before it starts. If it still can't run, you get a
503withRetry-After. - No uptime guarantee during the beta.
- One model. If you need another one, you need another provider.
How this compares with every other public offer for the same model: the price comparison.