Auto cascade

sieve-auto gives you 31B accuracy on hard questions for close to the price of the small model.

1. Every question is first answered by Sieve-4B.
2. A question whose confidence is below its threshold is asked again on Sieve-31B.
3. The two answers are merged into one response.

Only the uncertain questions go to 31B. Confident answers stay on 4B.

curl https://api.sieve.one/v1/systemone \
  -H "Authorization: Bearer $TYPESAFE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "model": "sieve-auto", "state": "...", "questions": { "q": { "type": "noul", "instructions": "..." } } }'

ThresholdsCopy Anchor Link

Each question type has its own confidence threshold. A higher threshold sends more questions to 31B. For noul, which has no public confidence, the cascade uses an internal measure of how far the probability is from 0.5.

Today every threshold is 0.5. These are placeholder values until they are tuned per question type, and they may change.

Which model answeredCopy Anchor Link

The response's model field names the largest model that answered: sieve-31b-v1 when at least one question was asked again, otherwise sieve-4b-v1. To see which model answered each question, send the header Sieve-Extras: true. The response then has a top-level sieve object with answered_by (model per question name) and escalated. This works on System One and /v1/decisions. See Migrate from Jev.

PriceCopy Anchor Link

• Modelled list price: $0.050 per 1M input tokens, output free.
• For comparison, Sieve-4B is modelled at $0.025, Sieve-9B at $0.042 and Sieve-31B at $0.160 per 1M input tokens.
• These prices are modelled from GPU costs. Final prices come after the fan-out benchmark. See Pricing.
• An Auto request is billed at the Auto price for its input tokens, whether or not any question escalates. Re-asked questions are not billed again at the 31B price.

When to use itCopy Anchor Link

• Use sieve-auto when most questions are easy and a few are hard.
• Use sieve-4b or sieve-9b when you need the lowest cost and can accept lower accuracy on hard cases.
• Use sieve-31b when every question needs the highest accuracy.

Pin a versionCopy Anchor Link

Aliases such as sieve-auto and sieve-latest can move when a new version ships. To pin Auto, send its versioned name, sieve-auto-v1. Do not pin the model value an Auto response reports (sieve-4b-v1 or sieve-31b-v1): that calls the single model with no cascade. A versioned ID never changes behaviour.

Next stepsCopy Anchor Link