Home → Use cases

Use cases

Written from billing data on this platform, not from illustrative estimates.

Coding agents

Claude Code, Codex, Cursor and Cline against one OpenAI-compatible endpoint. What drives the bill in an agent loop, and the base-URL detail that waste

RAG and retrieval pipelines

Where the cost sits in a retrieval pipeline, why the generation model is rarely the expensive part, and how to split cheap and expensive steps across

Support and chat bots

How to keep a support bot's cost predictable, why output length matters more than model tier, and what to check before pointing one at customers.

What these have in common

Three different workloads, one pattern: the model tier is rarely the biggest lever on the bill. In agent loops it is cache direction. In retrieval pipelines it is which stage gets the expensive model. In support bots it is output length. Each page works through the arithmetic for that specific shape of traffic, using billing data from this platform rather than illustrative estimates.

Setup is the same for all of them

One OpenAI-compatible base URL covers every model, so routing a stage to a cheaper model is a one-word change rather than a second integration. Use https://apiclan.us/v1 for OpenAI-lineage clients and SDKs, and https://apiclan.us with no /v1 for Claude Code and the Anthropic SDK. The base URL guide has the exact setting for each client and explains why a mistake there produces an error that appears in no log at all.

Keys are bound to one group. A key created for text models will be refused by image models, which is the single most common surprise for new accounts. Which models sit in which group is on the price list.

Measuring your own workload before you optimise anything

Every recommendation on these pages assumes you know the shape of your own traffic, and most teams do not until they look. Two numbers settle it, and both are in your usage history:

Cost per call, median and mean. If the mean is much higher than the median, a small number of large calls is carrying your bill and the fix is in those calls, not in the model choice. On our own traffic that ratio is roughly three to one.
The share of input arriving as cache reads. A low share in a workload that resends the same context means the cache is being rebuilt rather than reused, which is the most expensive way to run an agent and the cheapest thing to fix.

Optimising before measuring usually means changing the model, which is the slowest lever to validate and often the smallest.

What is not covered here

Embeddings, fine-tuning and batch endpoints are provider account-level features and do not pass through a relay. If your pipeline needs them, those calls go to the provider directly while generation and reasoning calls can still come through here.

Start using it

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 quickstart

Prices quoted on this page are regenerated automatically from live billing data. Third-party terms are quoted from that party's own published documentation.