Introducing Vana, a System One model

We are launching Vana, a System One model from Annamal.ai, built for decisions rather than text. You send it a state and typed questions. It answers every question in a single forward pass, with a probability and a confidence on each answer. There is nothing to parse. Vana is available now through the API.

What Vana does

A System One model is built for fast, structured judgement. Where a language model writes an answer one token at a time, Vana reads the situation and decides. Each question has one of three types:

• choice: pick one of 2 to 255 options. You get the option and a probability for each one.
• noul: yes or no. You get the probability of yes.
• score: place it on an ordered scale of 2 to 10 levels. You get the expected level and the distribution.

Every answer carries a confidence, so your code can decide when to act and when to ask a person. You can put up to 64 questions in one request, and the time stays about the same however many you ask.

curl https://api.vana.annamal.ai/v1/systemone \
  -H "Authorization: Bearer $VANA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "vana-latest",
    "state": "Customer: I was charged twice for my October invoice, please fix this today.",
    "questions": {
      "refund":  {"type": "noul",   "instructions": "Should we issue a refund?"},
      "team":    {"type": "choice", "instructions": "Which team should handle this?",
                  "criteria": {"billing": "Billing / payments", "tech": "Technical issue", "sales": "Sales"}},
      "urgency": {"type": "score",  "instructions": "How urgent is this?",
                  "criteria": ["not urgent", "normal", "urgent", "critical"]}
    }
  }'

{
  "model": "vana-1.0",
  "answers": {
    "refund":  {"type": "noul", "noul": 0.93, "probabilities": {"true": 0.93, "false": 0.07}, "confidence": 0.86},
    "team":    {"type": "choice", "choice": "billing", "probabilities": {"billing": 0.91, "tech": 0.06, "sales": 0.03}, "confidence": 0.91},
    "urgency": {"type": "score", "score": 2.6, "legend": {"1": "not urgent", "2": "normal", "3": "urgent", "4": "critical"},
                "probabilities": {"1": 0.02, "2": 0.38, "3": 0.52, "4": 0.08}, "confidence": 0.52}
  },
  "usage": {"input_tokens": 112, "output_tokens": 3}
}

The benchmark

We wrote one realistic support message and 27 typed questions about it: 14 yes or no, 7 choice and 6 score, covering billing, technical, security, urgency and churn signals.

Customer (Acme Corp, Enterprise plan): Our integration has been throwing 500 errors since 9am and two other accounts on our team are seeing the same thing. This is blocking a product launch scheduled for tomorrow with a major partner. We also noticed we were charged twice on last month's invoice. If this isn't fixed today we are going to have to consider other vendors.

Vana received the 27 questions in one request. GPT-OSS-20B received the same message with a system prompt that listed every field and its allowed values and asked for a single JSON object. Temperature 0, streamed. Jev takes the same request as Vana, so we priced it on the same input.

"questions": {
  "revenue_impacted":   {"type": "noul",   "instructions": "Is revenue currently impacted?"},
  "business_impact":    {"type": "choice", "instructions": "What is the business impact?",
                         "criteria": {"none": "No business impact", "minor": "Minor inconvenience",
                                      "degraded": "Degraded but usable", "severe": "Severe, blocking"}},
  "churn_likelihood":   {"type": "score",  "instructions": "How likely is the customer to churn?",
                         "criteria": ["1", "2", "3", "4"]},
  ...24 more
}

* Jev is priced on the same 171 input tokens at its published rate. Its own token count may differ.

How we measured

• Latency figures are averages from our tests.
• Cost uses each model's published price: $0.035 per 1M input tokens for Vana and $0.042 for Jev, with output free on both. For the LLM we used the cheapest public price we found for GPT-OSS-20B, $0.018 per 1M input and $0.09 per 1M output tokens.
• The LLM ran on a free preview endpoint. Results will differ on other hosting, and with a frontier model.
• It is one incident. Run your own messages and questions to see your numbers. Jev is faster on average, and reads much longer inputs, up to 64k tokens against Vana's 1,024.

Price and limits

Vana is prepaid, with no subscription. You pay $0.035 per 1M input tokens and output is free. Top up from $5. Vana 1.0 reads English text, either a string or JSON, up to 1,024 input tokens per request, and keys are limited to 300 requests per minute by default.

What to build with it

• Support ticket routing. Send each ticket to billing, tech or sales.
• Churn-risk detection. The probability that a message signals intent to cancel.
• Urgency scoring. Rate requests from not urgent to critical.
• Content moderation. Flag policy violations and pick the category.
• Intent classification. Map an utterance to one of your intents.
• Lead qualification. Score inbound leads against your own criteria.

Get started

Create a key, top up $5 and send one request. The quickstart takes about two minutes.