Use Otium behind a gateway

A gateway reaches Otium through the same OpenAI Files and Batch API your own code would use, so batch work passes through unchanged. Otium has no synchronous chat endpoint, on purpose: it is where work goes when nobody is waiting on the answer.

Otari is Mozilla AI's open-source gateway. Add Otium as a named instance in its config.yml, then call models as otium:<model>:

providers:
  otium:
    provider_type: openai
    api_base: "https://api.getotium.ai/v1"
    api_key: ${OTIUM_API_KEY}
    models: [otium-small, otium-medium]

# Otium's Standard rates, for batches sent with completion_window "24h".
pricing:
  otium:otium-small:
    input_price_per_million: 0.046
    output_price_per_million: 0.23
    cache_read_price_per_million: 0
  otium:otium-medium:
    input_price_per_million: 0.80
    output_price_per_million: 4.00
    cache_read_price_per_million: 0

Otari prices requests from its own config, so these are copies of the rate card as of 2026-10-19. They include a cached-input price of zero, because input served from cache is free on Otium. Without it, Otari would charge cached tokens at the full input rate and show more than we bill.

Otari keeps one price per model, so the config above assumes its default window, 24h (Standard). If you send 7d batches (Remnant), use these prices instead:

providers:
  otium:
    provider_type: openai
    api_base: "https://api.getotium.ai/v1"
    api_key: ${OTIUM_API_KEY}
    models: [otium-small, otium-medium]

# Otium's Remnant rates, for batches sent with completion_window "7d".
pricing:
  otium:otium-small:
    input_price_per_million: 0.0345
    output_price_per_million: 0.1725
    cache_read_price_per_million: 0
  otium:otium-medium:
    input_price_per_million: 0.60
    output_price_per_million: 3.00
    cache_read_price_per_million: 0

Use Otari's batch endpoint. Otari writes the input file, uploads it to Otium and creates the batch. completion_window is the deadline, and it sets the price: 24h bills at Standard and 7d at Remnant.

curl https://your-otari.example.com/api/v1/batches \
  -H "Authorization: Bearer $OTARI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "otium:otium-small",
    "completion_window": "24h",
    "metadata": {"team": "catalog", "feature": "enrichment"},
    "requests": [
      {"custom_id": "sku-1", "body": {"messages": [{"role": "user", "content": "Tag this product: ..."}]}},
      {"custom_id": "sku-2", "body": {"messages": [{"role": "user", "content": "Tag this product: ..."}]}}
    ]
  }'

Up to 16 metadata pairs are yours. Otari adds otari_user_id to every batch to check who owns it; that key doesn't count toward your 16, and it comes back unchanged when Otari retrieves the batch.

• Failed requests. Like OpenAI, Otium writes failed requests to a separate error file. Otari's results read only the output file, so a request that failed is missing from them rather than shown as an error. Fetch the batch's error_file_id from Otium to see why it failed.
• The bill of record is ours. Today Otari prices a batch's results by the model each result line names (the model that served it, such as gemma-4-12b-qat-w4a16) under its OpenAI implementation, not by the otium: prices above, so its cost for a batch can be missing or wrong. Your Otium portal shows what was charged.
• A sync request says why. If a request reaches Otium on Otari's chat path, it fails with a 404 and code batch_only, and the message names the batch endpoints. Nothing is charged.

Any gateway that can send OpenAI batch calls to a custom base URL works the same way: https://api.getotium.ai/v1, an Otium API key, and the Files and Batch endpoints in the API reference. A gateway that only forwards synchronous chat can't send work to Otium.