One credit per turn, and what a turn is
No token counting, no per-tool fees, no charge for a decline. A plain explanation of how Bluwing pricing maps to the work the engine does.
Token-based pricing is a poor fit for tool calling. The customer cannot see the tokens, cannot predict them, and pays for a long tool list even when the request uses one function. The invoice is a measure of the prompt, not of the work.
Bluwing prices the work: one credit per turn that resolves to a tool call.
What counts as a turn
A turn is one request from your application that ends in a successful resolve. That means:
- the engine chose a tool;
- every required argument was grounded;
- a valid
tool_callsresponse was produced.
A turn that returns no_tool is free. A turn that returns needs_input is free; the user's answer produces a new turn, and only that one can be charged. Retries caused by the engine are free.
What a turn does not depend on
The price of a turn is the same regardless of:
- how many tools are in the tree;
- how long the conversation is;
- whether the path was fast or reasoning;
- which output format you asked for, OpenAI or Anthropic.
That last point matters more than it sounds. Token pricing makes a long context expensive, so teams trim history to save money and lose the facts the engine needs for grounding. Credit pricing removes that pressure: keep the conversation, keep the tool list, and pay for the calls you make.
Builds and edits
A tree is not a turn. Compiling a new one is charged once, by size:
| Action | Credits |
|---|---|
| New tree, up to 20 tools | 20 |
| Each additional tool in the build | 1 |
| Editing one tool | 1 |
Re-publishing a version does not re-charge the build if nothing changed. Aliases are free to move; they are pointers, not copies.
Plans and top-ups
The free tier includes 200 credits once, which is enough to test a real integration. Paid plans start at $4.50 a month for 4,500 credits, and the per-credit price falls as the plan grows. Top-ups never expire, and plan credits roll over for 60 days.
There are no per-second or weekly rate quotas. Your balance and your in-flight limit set the pace, so a burst of traffic is a billing event, not a 429.
A worked example
A support agent handles 3,000 turns a month. Each turn is one request, and roughly one in ten ends in no_tool because the user asked something outside the tool set. The billed turns are 2,700.
At Plan A's effective $0.001 per credit, that is about $2.70 of usage on a $4.50 plan. The remaining credits roll over. The same workload priced by tokens would depend on how many tools were in the context on every one of those 3,000 turns — the number the customer can least control.
Where the model fits
Bluwing does not charge separately for a model, because the fast path does not call one. On the reasoning path the engine may use one internally, and that cost is inside the credit. You are buying a result, not a chain of inference.
If you also run a plain model in your app for the no_tool case, that is billed by your model provider as usual. Bluwing charges nothing for the turn it declined.
Frequently asked questions
What happens when my credits run out?
Requests fail with a clear billing error; they do not silently degrade. Top up and the same request succeeds.
Do failed builds cost credits?
A build that fails validation does not consume its credits. A successful build is charged once, at the size it ended up.
Are credits shared across trees?
Yes. The balance belongs to the account, not to a tree. Every tree on the account draws from the same pool.