Batch pipelines at the lowest cost per token

Built for the overnight run

Batch is the workload where price per token is the whole decision. Enriching a CRM with a summary per account, classifying a year of support tickets, generating synthetic training data by the tens of millions of rows: none of it is latency-sensitive, all of it is volume. On the gpt-oss-120b class that volume costs $0.07 per million tokens blended, and gpt-oss-20b is the fastest stream in the catalog when the job leans small.

Utilization, not markup

The economics are fleet economics. Nyx serves more than 40 billion tokens a month, and batch traffic fills the troughs between interactive peaks, so the GPUs that price the fleet are busy around the clock. That is where 2.9× performance per dollar against list-price clouds comes from: hardware that never idles, an engine tuned per model, and a prefix cache that removes repeated prefill when every request in the job shares the same instruction block.

Per-token beats committed spend

Batch jobs are bursty. A committed-spend contract prices your peak night and bills you for the quiet ones; per-token pricing with no minimums bills the tokens the job used and nothing else. Capacity behavior is explicit too: past declared capacity Nyx returns 429 immediately with a Retry-After header, so a pipeline backs off on schedule instead of hanging workers against a queue. Run a thousand rows to validate the prompt, then run fifty million on the same endpoint at the same price.

# one worker in tonight's classification run
curl https://api.nyxprovider.com/v1/chat/completions \
  -H "Authorization: Bearer $NYX_API_KEY" \
  -d '{
    "model": "gpt-oss-120b",
    "messages": [{"role": "user",
      "content": "Label this ticket: refund, bug,
        or sales. Reply with the label only."}]
  }'

Recommended models