Switch from OpenRouter in one line.

Most OpenRouter ids work as sent. A Claude id written with a dot is read as ours: anthropic/claude-sonnet-4.5 runs as anthropic/claude-sonnet-4-5. A short list of other OpenRouter names map to the same release here, such as deepseek/deepseek-chat to deepseek-ai/deepseek-v3.

All seven fields are read: order, only, ignore, allow_fallbacks, sort (price, throughput or latency, as a string or { by }), max_price (prompt and completion, in dollars per million tokens) and require_parameters.

:floor sorts routes by price, :nitro by throughput, :online runs a web search first, and :free uses the free tier on models that offer it.

Server-sent events in the OpenAI chunk shape. The last chunk carries usage with its cost, without asking for it through stream_options.

Passed to OpenAI-compatible providers as sent, and translated to Anthropic's tool format on Claude routes.

json_object and json_schema. On Claude routes it runs as a forced tool call, so JSON mode works there too. The response-healing plugin repairs JSON that fails to parse.

The file content part, with file_data or file_url, on every chat model: Google and Anthropic models read the PDF itself, the others get its extracted text.

Middle-out drops messages from the middle of an over-long conversation so it fits the context window. The web plugin runs a search before the model answers; each search is billed.

Recorded with every request. An app that sends both from a public domain, with at least 10 requests in 30 days, is listed on our Apps page.

Every chat completion names the provider that served it and carries usage.cost in dollars./generation returns total_cost and the token counts of a past request.

pricing.prompt and pricing.completion as dollars per token, and context_length, next to our own per-million fields.

Not read: send one model per request and drop both fields. When a provider fails, the request already moves to the next provider serving the same model. A preset can name backup models; the first one with a working route is used.

Dropped without effect. Zero data retention, and whether providers that may train on prompts can serve you, are account settings here, under Settings, Data & privacy, and apply to every request.

Take the provider names on our Providers page. A name we do not route to matches nothing, so an only list made of those leaves no route and the request answers 503.

600+ models, fewer than OpenRouter lists at the time of writing. An id with no same-version model here answers 404 model_not_found; we never swap in a neighbouring version. A model we do not carry can be asked for through support.

Models that cost us nothing upstream are free to you: 50 requests a day per account, 1,000 a day once your paid usage reaches $10. Everything else, including older ids that end in :free, is billed at the provider price. :free on a model with no free route answers 400 model_not_free.

Tokens are billed at the provider list price, the same price per token as on OpenRouter. Our platform fee of 5% is charged once, when you buy credits, plus a 4% currency margin when you pay in rupees.

Work here too: add them in your dashboard. A request served on your key is billed 5% of what it would have cost on ours.

Read the same way, but a preset belongs to an account: create yours here before you switch.

Answers in our own shape, balance_usd and the latest transactions, so a script that reads your balance needs its field names changed.

The body carries { "error": { "code", "message" } } with a word for the code, such as model_not_found, and the number on the HTTP status. The OpenAI SDKs raise them as usual.

web and response-healing only; any other plugin id answers 400.