ToothFairyAI โ Decision Models API (Drex)
A decision model is not a chat model. It does not write prose: you give it the facts of a case (state) and one or more typed questions, and it scores every option you supplied in a single forward pass. The answer is a calibrated probability per option โ nothing else can come back.
Use it when you want a number you can compare against a threshold, not a sentence you must parse.
Decision models are intentionally excluded from /models_list and /v1/models, so they can never be selected as a chat model in the app or in TF Code. The authoritative list lives in the Decisions API reference and here.
Available modelsโ
| Alias | Size ยท context | Modalities | Best for |
|---|---|---|---|
drex-max | 27B ยท 262K | text, JSON, images (โค 8) | Flagship general decisions; computer-use screenshots |
drex-max-finance | 27B ยท 262K | text, JSON, images (โค 8) | Banking, accounting, payments, insurance, credit, compliance, filings, markets |
Endpointโ
POST https://ais.toothfairyai.com/decisions
POST https://ais.toothfairyai.com/v1/decisions # OpenAI-style path alias
Auth is the same workspace key used by every other inference endpoint โ x-api-key: <key> or Authorization: Bearer <key>. The workspace is resolved from the key, so no workspaceid body field is needed.
Question typesโ
type | You send | You get back |
|---|---|---|
noul | A yes/no question | noul = P(true) in [0,1] |
choice | criteria: a map of {option: description} | choice, confidence, probabilities{option: p} |
score | criteria: an ordered array of levels (โฅ 2) | score (expected level index), confidence, probabilities{"0": p, โฆ}, legend{"0": level} |
Several questions about one state share one pass over that state: asking three questions costs one decision, not three.
Quickstartโ
curl -s -X POST "https://ais.toothfairyai.com/decisions" \
-H "x-api-key: $TF_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "drex-max",
"state": "Invoice #4471 for 12,340 EUR, due 2026-10-01, paid 2026-10-14.",
"questions": {
"category": {
"type": "choice",
"instructions": "Classify the payment event",
"criteria": {
"paid_on_time": "Payment received on or before the due date",
"paid_late": "Payment received after the due date",
"unpaid": "No payment received"
}
},
"risk": {
"type": "score",
"instructions": "How risky is this account?",
"criteria": ["very_low", "low", "medium", "high", "very_high"]
},
"fraud": {"type": "noul", "instructions": "Is there any indication of fraud?"}
}
}' | python3 -m json.tool
{
"id": "decision-9f2b1c0a4d5e6f708192a3b4",
"object": "decision",
"created": 1763412789,
"model": "drex-max",
"answers": {
"category": {
"type": "choice",
"choice": "paid_late",
"confidence": 0.99995,
"probabilities": {
"paid_on_time": 0.00002,
"paid_late": 0.99996,
"unpaid": 0.00002
}
},
"risk": {
"type": "score",
"score": 1.55,
"confidence": 0.44,
"probabilities": {"0": 0.057, "1": 0.438, "2": 0.412, "3": 0.082, "4": 0.011},
"legend": {"0": "very_low", "1": "low", "2": "medium", "3": "high", "4": "very_high"}
},
"fraud": {"type": "noul", "noul": 0.103}
},
"usage": {
"prompt_tokens": 323,
"completion_tokens": 374,
"total_tokens": 697,
"prompt_tokens_details": {"cached_tokens": 0},
"completion_tokens_details": {"reasoning_tokens": 0},
"uoi": 1
},
"latency_ms": 39.58
}
Imagesโ
Pass up to 8 images in images as data URLs (data:image/png;base64,โฆ) or public https:// URLs. Images are billed as input tokens and reported separately as usage.prompt_tokens_details.image_tokens, already included in prompt_tokens.
{
"model": "drex-max",
"state": "What should the agent click next?",
"images": ["data:image/png;base64,iVBORw0KGgo..."],
"questions": {
"next": {
"type": "choice",
"instructions": "Pick the next action",
"criteria": {"click_checkout": "The checkout button", "click_cart": "The cart icon"}
}
}
}
Thresholding the answerโ
The value of a decision model is the calibrated probability, not the argmax. The usual pattern is:
answer = resp["answers"]["fraud"]
if answer["noul"] >= 0.9:
auto_reject()
elif answer["noul"] <= 0.1:
auto_approve()
else:
send_to_human_review() # the model says it is unsure โ route it
Region โ global onlyโ
Decision models run on a single global pool. There is no regional pool, no cross-region failover and no data-residency guarantee.
region therefore accepts only null (just omit it) or "global". Any other value โ "au", "us", "eu" โ is rejected with a 400 rather than quietly served from the global pool, so you can never believe you received regional routing:
{"error": {"message": "Decision models are global-only; region 'eu' is not available. Omit 'region' or send 'global'.", "type": "invalid_request_error", "code": 400}}
Billingโ
- Billed per request in UoI (1 UoI = $0.01), reported as
usage.uoi. - Decision models are metered on input tokens; output is never billed.
- A typical decision costs a fraction of a cent, so each request is floored at 1 UoI.
usage.prompt_tokensis the exact count (image tokens included). There is no token-buffer markup on this path.- Workspace rate limits and the monthly UoI budget apply exactly as on
/predictions.
Strict request schemaโ
Unknown fields are rejected with a 400 that names the field โ including every chat-shaped parameter (messages, stream, tools, tool_choice, temperature, top_p, top_k, max_tokens, stop, reasoning_effort, response_format). A decision request has none of them.
| Situation | Response |
|---|---|
Chat alias sent to /decisions (e.g. sorcerer) | 400 pointing at /predictions |
Decision alias sent to /predictions | 400 pointing at /decisions |
Invalid question definition (e.g. choice with a list criteria) | 400, message names the question |
| More than 8 images / inline payload > 10MB | 400 |
Any region other than global (au / us / eu) | 400 โ decision models are global-only |
Further readingโ
- Decisions API reference โ full request/response schema
- Models & Regional Endpoints โ chat models (decision models are listed here, not there)
- Units of Intelligence โ how UoI consumption is calculated