Aider on one prepaid key

• Aider installed (aider on your PATH) and a git repository to work in.
• A RelaySwitch key. Sign in with an email code at the console and create one. Keys start with sk-relay-.
• A balance. Top-ups start at $10, by card or crypto, pay as you go. Before the first top-up the console chat gives you 12 free turns a day to try models; aider requests bill the balance.

Aider's OpenAI-compatible API page (read 2026-10-28) asks for two environment variables and the openai/ prefix on the model name. The prefix tells aider which client to use; the id after it is the one from the Models page.

On Windows the docs use setx and a fresh shell:

Keep the /v1 on the base URL. Aider adds /chat/completions itself. To switch models, change the id after the prefix; the base URL and key stay.

Aider reads .aider.conf.yml from your home directory, then the git root, then the current directory, and the last file loaded wins (config docs, read 2026-10-28). The YAML keys are the long flag names without the dashes in front:

The same settings as a .env file, searched in the same order (.env docs, read 2026-10-28). Aider's own options take an AIDER_ prefix; the key stays the plain OPENAI_API_KEY the client reads:

A file in the home directory covers every project. A file at a git root pins one project to another model. Keep the key out of any file you commit: a .env at the git root belongs in .gitignore.

weak-model is the model aider uses for commit messages and chat-history summaries. Those calls are short and frequent, so a cheap one does; Claude Haiku 5.5 bills $0.0582 in and $0.1455 out per million tokens today. Without the line, aider sends those calls to the main model and bills them at its price.

For a model it has not heard of, aider prints "Unknown context window size and costs, using sane defaults", treats the window as unlimited and the model as free (warnings page, read 2026-10-28). The fix is a .aider.model.metadata.json in the home directory, the git root or the current directory (advanced model settings, read 2026-10-28). One entry per model, under the prefixed name, with costs in USD per token:

The numbers above come from the live catalog when this page was built: the context window as max_input_tokens, the output ceiling as max_output_tokens, and the per-million prices divided by a million. They feed aider's token counts and the cost line after each reply. Billing does not read them; every request bills at the catalog price on the console's Usage page.

Two curl calls rule out the key and the base URL before you read aider's output. The first lists every model id your key can call:

The second sends one small request the way aider does. -i prints the headers, including the request id and, on a non-streamed answer, the charge in USD.

If both work, any failure left is in aider's settings. Send a message in aider and the request shows on the console's Usage page within seconds, with its tokens and its charge.

Aider resends the repo map, the files you added and the chat history with every message, so input tokens are most of the bill. A low input price matters more than a low output price. Of the models here, Kimi K2.7 Code has the lowest input rate today at $0.426 per million tokens.

Claude ids answer the OpenAI format, so Claude Sonnet and Opus run under the openai/ prefix with no Anthropic variable. For the full list, with every price and context size, see the Models page. How the balance and the cached rate work is on Pricing.

The session: one feature built over 40 aider messages, late enough that the chat carries a large repo map and several files. Each message sends 80,000 tokens of earlier context, 2,000 new input tokens, and gets 1,500 output tokens back. Those sizes are an assumption; your own Usage page shows the real ones.

Two columns, because whether the earlier context hits the prompt cache depends on the model and the request. When it does, the response reports cached tokens and they bill at the cached rate. The console's Usage page shows the split for every request.

With cache hits the same session runs $0.3346 on DeepSeek V4 Pro 0813 and $1.26 on Claude Opus 5.5. Aider's /model command switches the main model mid-chat, so a cheap model for routine edits and a dear one for the hard step costs nothing to set up.

• 401 or an authentication error from the client: the key is wrong, revoked or cut short. Paste it again from the console and run the curl check.
• 402: the balance is too low for the request. Top up in the console; the key works again at once.
• 404 or a model-not-found error: the id after openai/ is not in the catalog. Check for a typo or a display name in place of the id; GET /v1/models lists every id.
• 429: the key hit its rate limit. Wait a minute and resend, or create a key with a higher Requests / min in the console.
• Connection errors: OPENAI_API_BASE is missing /v1, has a typo, or a proxy blocks it. The curl check from the same machine tells you which.
• "Unknown context window size and costs": aider has no metadata for the id. Add the metadata file above; the model works either way.
• Aider asks for a variable the model needs and names one you have not set: the model id is missing the openai/ prefix, so aider looked for another provider's key.