Vana, a System One model from Annamal.ai

System One Model
for decisions.

Send a state and typed questions to one endpoint. Get choices, probabilities and scores back from a single forward pass. It never writes text, so there is nothing to parse.

Get API key Quickstart

POST /v1/systemoneEnglishabout 200 ms server-sideprepaid from $5

Request · curl

Response

Keys start with vana_live_ or vana_test_. Call Vana from your backend, never from browser or mobile code.

Full quickstart API reference SDKs

01 · System One

Judgement in one pass,
not text one word at a time.

Vana is a System One model: the fast layer of an AI stack that reads a situation and judges it. A language model stays the slow layer that writes and reasons. Vana does not generate text. It answers your typed questions all at once.

endpoint: POST https://api.vana.annamal.ai/v1/systemone
model: Vana 1.0 (vana-1.0). The aliases vana-1 and vana-latest follow new versions.
request: A state (text or JSON, English, up to 1,024 input tokens) and up to 64 typed questions keyed by your own ids.
question types: choice, noul (yes or no) and score.
answer: The choice, probability or score, with probabilities over the options and a confidence. No text to parse.
latency: About 200 ms server-side, and about the same for one question or sixty-four, because they are answered in one forward pass.
billing: 1M input tokens free on signup, then $0.035 per 1M. Output is free. Prepaid credit from $5, no subscription.

02 · Compare

Vana and an LLM, side by side

Both get the same situation and the same twelve questions. Press Run to see how each one answers.

Situation

Show answers as

Same input for both

Vana

LLM

Press Run to start.

Simulated. Sample situations use sample data; text you write is answered in preview mode. Timings are illustrative, not benchmarks.

03 · Vana vs Jev

Already on Jev?
Moving takes three lines.

Jev, TypeSafe AI's System One model, takes almost the same request as Vana: same path, same JSON body, same three question types. For plain-text questions, moving takes three lines. Vana is 17% cheaper per input token. Jev is faster and reads much longer inputs. Here is everything that differs, so you can decide.

Input price per 1M tokens

$0.035Vana

$0.042Jev

Vana is 17% cheaper. Output tokens are free on both.

Cost of 1M requests, 150 input tokens each

$5.25Vana

$6.30Jev

Vana saves $1.05 per million requests.

Average latency

0.200 sVana

0.114 sJev

Jev is faster. Both stay flat as you add questions.

	Vanaby Annamal.ai · `vana-1.0`	Jevby TypeSafe AI · `jev-1.13.0`
input price	$0.035 per 1M tokens 17% lower	$0.042 per 1M tokens
output price	free	free
1M requests, 150 input tokens	$5.25	$6.30
average latency	0.200 s	0.114 s faster
context per request	1,024 input tokens	64k tokens, 32k for the state plus the longest question longer
rate limit	300 requests per minute per key, more on request	100K tokens and 80 requests per second
input	English text: a string or JSON	text: a string, JSON object or array
endpoint	`POST api.vana.annamal.ai/v1/systemone`	`POST api.typesafe.ai/v1/systemone`
request body	`model`, `state`, `questions` with `type`, `instructions`, `criteria`. The same in both for plain-text questions. Jev also accepts structured `instructions` (an object or array).
score levels	count from 1: `"1"` to `"n"`	count from 0: `"0"` to `"n-1"`
how you pay	prepaid credit from $5, drawn down per call. No subscription.	charged per input token

Jev details from its models page, its API reference, the quickstart and TypeSafe's launch post, checked 13 October 2026. We have not run Jev ourselves. Their prices and limits can change. Latencies are averages from our own tests. We have not benchmarked accuracy against Jev, so test both on your own data.

The three lines that change, over HTTP

- curl https://api.typesafe.ai/v1/systemone \
+ curl https://api.vana.annamal.ai/v1/systemone \
-   -H "Authorization: Bearer $TYPESAFE_API_KEY" \
+   -H "Authorization: Bearer $VANA_API_KEY" \
    -H "Content-Type: application/json" \
    -d '{
-     "model": "jev-latest",
+     "model": "vana-latest",
      "state": "...",
      "questions": { ... }
    }'

Both SDKs take different calls (TypeSafeClient.system_one against Vana.decide), so if you use an SDK, swap it or call HTTP directly. See our SDK docs.

Check before you switch

Input length. Vana reads up to 1,024 input tokens and truncates the rest. Jev reads up to 64k. Short messages, tickets and records fit. Long documents do not.
Score indexes. Vana counts levels from 1, Jev from 0. Shift any code that reads a level number.
Answer fields. Read answers by name: choice, noul, score. Vana yes or no answers carry probabilities and confidence as well as noul. Jev's API reference lists only noul for them.
Structured instructions. Jev lets instructions be an object or array. Vana's are text, so flatten any structured instructions into a string.
Errors and limits. Jev returns 401, 422, 429 and 529. Vana adds 402 when your prepaid balance runs out and 403 for an inactive account, so handle those. Rate limits differ too. See our errors page.
Accuracy. Run both on a sample of your own data before moving traffic.

cost estimate, per monthedit the numbers

input tokens per request requests per month an LLM doing the same job LLM output tokens per request

vana-1.0$5.25

jev-1.13.0$6.30

llm$130

vana-1.0 costs 17% less than jev-1.13.0 here, and about 25 times less than the LLM.

Vana and Jev bill input tokens only. The LLM price range is the one TypeSafe publishes for LLMs, with output at about 5 times input. These are assumptions to edit, not our measurements.

04 · Question types

Three question types

Every question has a type, free-text instructions and type-specific criteria. Put the facts in the state and the decision in the instructions.

question

answer

State for all three: "Customer: I was charged twice for my October invoice, please fix this today." Reference

Choose how much to automate

Every answer comes with a confidence. Pick a number. Answers above it can run on their own, and the rest go to a person.

Automate above80%

0run automatically

0go to a person

Uses the situation selected in the comparison above.

05 · Use cases

What you can build

Decisions that come up again and again. These are examples, not customer claims.

Support ticket routingSend each ticket to billing, tech or sales.choice

Churn-risk detectionThe probability that a customer message signals intent to cancel.noul

Urgency scoringRate incoming requests from not urgent to critical.score

Content moderationFlag policy violations and pick the category.noul + choice

Intent classificationMap a user utterance to one of your intents.choice

Lead qualificationScore inbound leads against your own criteria.score

Built on Vana

Naaledge AI uses Vana for faster decisions.

Naaledge AI is a separate Annamal.ai product. Vana is the decision model underneath it.

Visit Naaledge AI

06 · Pricing

Pay per input token.

You buy credit up front and each call draws it down. Output tokens are free, and a call takes about the same time whether it carries one question or sixty-four.

rate: $0.035 per 1M input tokens
output: free
latency: about 200 ms, server-side
credit: top up $5, $20, $100 or any amount from $5. No subscription.
expiry: none while your account is active. 12 months after your last activity.
at $0: requests return 402 insufficient_balance until you top up

cost estimateedit the numbers

input tokens per request requests per

$5.25 / month150,000,000 input tokens. $5 covers about 952,380 requests.

Vana 1.0 reads up to 1,024 input tokens per request; longer input is truncated. Pricing is subject to change. Pricing details

07 · FAQ

Frequently asked questions

Not here? Read the docs or write to support@annamal.ai.

What is Vana?

Vana is a System One model from Annamal.ai, available as a typed API. You send a state (text or JSON) and typed questions. You get back typed, calibrated answers with probabilities and confidence, from one forward pass. The current model is Vana 1.0.

What is a System One model?

A model built for fast, intuitive judgement instead of writing text. It reads a situation and decides in one pass, and your code gets typed answers. Vana's endpoint is /v1/systemone. A language model stays the slow layer that writes and reasons.

How is it different from an LLM?

An LLM writes its answer one token at a time, so every extra question adds waiting, and you parse the text. Vana answers all questions in one forward pass and returns typed results with probabilities, so software can act on them directly.

How fast is it?

About 200 ms server-side per call, however many questions you ask in it. The comparison above is a simulation with illustrative timings, not a benchmark.

What does it cost?

$0.035 per 1M input tokens, and output is free. You buy credit from $5 and each call draws it down. There is no subscription and no seat pricing.

How is Vana different from Jev?

Jev is TypeSafe AI's System One model. It takes almost the same request as Vana: same path, same JSON body, same three question types. Vana costs 17% less per input token. Jev reads far longer inputs, up to 64k tokens against our 1,024. If your inputs are short messages, tickets or records, Vana costs less. If you need long documents, use Jev. See the full comparison above.

What can it read?

Vana 1.0 reads English text, either a string or structured JSON. Input longer than 1,024 tokens is truncated, and the response tells you how many tokens were processed.

Can I call it from a browser?

No. Keep your key on your backend or an edge function. Never ship a key in browser or mobile code.

How does Vana relate to Naaledge AI?

Naaledge AI is a separate Annamal.ai product that uses Vana to make faster decisions. Vana is the model. Naaledge is a product built on top of it.

Make your first decision call.

Get API key Read the docs

System One Modelfor decisions.

Judgement in one pass,not text one word at a time.