Home → Use cases → Coding agents
Agent loops are the workload where a per-token relay makes the most difference, because they resend enormous amounts of context and most of it is cacheable.
An agent re-reads your codebase on almost every turn. That means input tokens dominate the count, and how they are billed depends entirely on whether they arrive as a cache read or a cache write.
From one real session on this platform: 30 requests, $1.13 charged, and 63% of that was cache writes — context being rewritten, producing no output at all. Cache reads cost a fraction of the input rate; cache writes cost more than it. Same tokens, opposite directions.
The practical consequence is that the biggest lever is not the model tier. It is keeping the context stable between calls so the cache is read rather than rebuilt. Reordering a prompt so the stable part comes first, or avoiding a tool that rewrites the whole file list each turn, moves the invoice more than dropping a tier does — and it does not require re-validating any output.
The second lever is routing. Most loops fire the expensive model at every step, including classification and extraction steps a cheap model handles identically. Sending only the final reasoning step to a flagship usually cuts blended cost by more than half.
| Client | Base URL |
|---|---|
| Claude Code | https://apiclan.us |
| Codex, Cursor, Cline, Roo Code | https://apiclan.us/v1 |
The split is not arbitrary: Anthropic-lineage clients append
/v1/messages themselves, OpenAI-lineage clients append only
/chat/completions. Get it backwards and you receive a 404 from the edge,
which means the request never reaches the gateway and never appears in your usage
log — it looks exactly like a missing model or a bad key. Full settings per
client are in the base URL guide.
Long agent runs are what claude-fable-5-1 is built for, and its cache reads
cost a quarter of what they do on other Claude models — in a loop where most input
is cache reads, the model with the higher headline price can end up cheaper per session.
It rejects forced tool use with a 400, so keep tool_choice on
auto.
claude-sonnet-5 is the common default for interactive work. On our own
traffic the median call costs 0.0087 credits and the mean is 0.0255 — a threefold
spread on one model, entirely from cache behaviour. Prices for every model are on
the price list.
No subscription, no monthly minimum, no sales call. Top up with USDT and spend what you use — 1 USDT gives you 2 credits of API balance.
Read the 30-second quickstartPrices quoted on this page are regenerated automatically from live billing data. Third-party terms are quoted from that party's own published documentation.