Chat completions

101 language models from 10 makers on one OpenAI-compatible endpoint. Point any OpenAI SDK at Revana, change the model name, pay per token from the same prepaid credit as video and images.

Make a request

curl https://api.revana.si/v1/chat/completions \
  -H "Authorization: Bearer $REVANA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-5.5","messages":[{"role":"user","content":"Write a haiku about routing."}]}'

The answer is OpenAI’s chat.completion object. Its id starts chat_, and usage adds cost_usd, the amount taken from your balance for this request.

{
  "id": "chat_3f9c…",
  "object": "chat.completion",
  "created": 1793318400,
  "model": "gpt-5.5",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "Cheapest path first,\nfallback waits in the wings,\nanswers still arrive."
      },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 14,
    "completion_tokens": 18,
    "total_tokens": 32,
    "prompt_tokens_details": {
      "cached_tokens": 0
    },
    "cost_usd": 0.000141
  }
}

OpenAI SDKs

Every OpenAI client library works unchanged: set the base URL and the key.

import os
from openai import OpenAI

client = OpenAI(base_url="https://api.revana.si/v1", api_key=os.environ["REVANA_API_KEY"])
reply = client.chat.completions.create(
    model="gpt-5.5",
    messages=[{"role": "user", "content": "Write a haiku about routing."}],
)
print(reply.choices[0].message.content)

import OpenAI from "openai"

const client = new OpenAI({ baseURL: "https://api.revana.si/v1", apiKey: process.env.REVANA_API_KEY })
const stream = await client.chat.completions.create({
  model: "gpt-5.5",
  messages: [{ role: "user", content: "Write a haiku about routing." }],
  stream: true,
})
for await (const chunk of stream) process.stdout.write(chunk.choices[0]?.delta?.content ?? "")

Tools such as LangChain, LlamaIndex, the Vercel AI SDK, Cline or Open WebUI take the same two settings: an OpenAI-compatible provider with base URL https://api.revana.si/v1 and your key.

Streaming

Send "stream": true for server-sent events in OpenAI’s chat.completion.chunk format, ending with data: [DONE]. Add "stream_options": {"include_usage": true} for a last chunk with empty choices and the request’s usage, cost_usd included.

curl https://api.revana.si/v1/chat/completions \
  -H "Authorization: Bearer $REVANA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-5.5","messages":[{"role":"user","content":"Write a haiku about routing."}],"stream":true}'

Billing

• You pay for the tokens the model reports: input, cached input, cache writes and output (reasoning tokens count as output), at the model’s price per million below. Charged to the micro-dollar from the same prepaid balance as generations.
• While a request runs, Revana holds what it could cost: its prompt plus max_tokens (or max_completion_tokens; 8,192 when neither is set). Set max_tokens to hold less. The hold is released when the request ends and only the tokens used are charged.
• A request with no answer, because every route failed, costs nothing. A stream you close part way is charged for what was sent.
• Some models price by prompt size, time of day or reasoning effort; each model’s page says so. Prompt caching works as at the model’s maker, and cached tokens are billed at the cached rate.
• Every model has a route at its listed price. When that route is busy, a request can run on a backup route, and is charged that route’s cost where it is higher, never more than the hold.

Errors

Errors use OpenAI’s shape, {"error": {"message", "type", "code"}}, so SDKs raise their usual exceptions: 400 invalid_request, 401, 402 insufficient_credits (top up, or lower max_tokens), 404 model_not_found, 413 body_too_large (over 20 MB), 429 concurrency_limit, 502 model_unavailable and 503 model_busy | busy (retry in a few seconds; nothing was charged).

Models and prices

USD per million tokens. Click a model for its full price sheet. The same list is /data/llms.json.