What is a decision model, and how is it different from a chat model?

A chat model writes its answer token by token, so a short answer takes half a second to several seconds and rarely answers the same way twice. DecisionNode reads your input once and answers your typed questions in one pass, each answer with how sure the model is: a choice, a score, a truth or a number. It generates no text, so it answers in milliseconds. A System One model is a class of AI model that returns typed, probabilistic decisions from a fixed answer space instead of generating text. Given a piece of state and a set of questions defined in code, it answers each one with a schema-valid value such as a chosen option, a score, or a probability, and then stops. That is the right tool for routing, triage, moderation, scoring and gating, and the wrong one for writing.

What are the question types?

Every type names its answer: choice (one of your labels), score (a level on your scale), truth (a probability), number (a value on your grid), point (a spot in your image), box (a rectangle around the thing). Choice picks one option from the list you send and returns a probability for each. Score places the input on your ordered scale, for example routine to critical, and returns the expected level. Truth, truth in the API, asks whether a statement is true: one calibrated probability from 0.00 to 1.00, so 0.80 means true about 8 times in 10. Your code picks the cut-off: 0.50 for a plain yes or no, higher when acting on a false yes is costly. Number, number in the API, asks: How many, or what value? One value on a grid you set (min, max and an optional step), with a calibrated probability for every value: the most probable value, the expected value and the confidence. A count of cars in a photo, overdue invoices in a statement, a year, a rating out of ten. Point, point in the API, and box, box, answer where. A point is one spot in the image, as x and y, with the probability that the thing is there at all (present) and that the spot is on it (confidence). A box is a rectangle around the thing, as its top-left corner (x1, y1) and bottom-right corner (x2, y2), with present and confidence.

How does pricing work, and why is output free?

You pay per million input tokens from a prepaid balance: $0.042 on DecisionNode MoE, $0.035 on DecisionNode Flash. Image and video tokens are priced per million too, $0.09 on DecisionNode MoE and $0.08 on DecisionNode Flash, with no fee per image. Work that can wait runs as a batch, 30% off the live price on DecisionNode MoE and 50% off on DecisionNode Flash. The model writes no text, so every response reports output_tokens: 0 and output never costs anything. No seats, no minimums. A repeated state is billed in full each time, with no discount for input sent before. Your first workspace has a free tier: 50 DecisionNode Flash requests, 20 DecisionNode MoE requests and 10 booster calls every day, from the playground and from API keys alike.

Which model should I pick, DecisionNode MoE or DecisionNode Flash?

DecisionNode Flash when speed and volume come first, at $0.035 per million input tokens. DecisionNode MoE for the hard calls, where accuracy matters most. Many builds run both: DecisionNode Flash first, and any answer under your confidence threshold is re-checked on DecisionNode MoE before your code acts.

Yes. Send up to 16 images next to your text in the same request, as base64 or an https URL. The model reads printed and handwritten text, amounts, dates, objects and layout, and answers through the same question types: count the cars in a camera frame, check a parcel on the belt, match a receipt to a claim.

Can it say where something is in an image?

Yes. Two question types, point and box, answer where. Ask "Point to the yellow package closest to the robot arm" or "Draw a box around the person at the door" and get coordinates you can draw, crop or hand to a machine. Every answer also says how likely it is that the thing is there at all (present), so your code acts on a sure answer and takes the safe path on an unsure one. Each place carries a confidence too, not calibrated yet, so act on present. The instructions can carry conditions and refer to the state and to other images, which a fixed list of detector classes cannot: production lines and sorting, quality control, surveillance and more. Billed by the rows it runs, like any question: the input tokens of its presence row and of the coordinate rows of the place it answers (two for a point, four for a box), about 140 tokens for a point and 250 for a box, plus the image once per request at the image and video rate. No fee per point or box, and output is free.

Can it watch traffic, people or a production line?

Yes, from camera images and from video sent as frames. Monitoring areas comes first: traffic and its flow (counts, turns, congestion, wrong-way drivers, pedestrians crossing), people in an area (counts, flow, crowding) and incidents (break-ins, theft, fire and smoke). Then production lines and sorting: one decision per item at the sorter (damaged, mislabelled, upside down, which lane), counting products, spotting jams and stoppages. Drones can be monitored from above through their camera frames, and a session hands a drone's own controller a typed decision for every frame. Warehouse machines get command-level decisions the same way: go, stop, which aisle. Need it on your own hardware, on premises, at the edge or next to the cameras? For qualifying deployments, talk to us at hello@bynn.com.

Yes, sent as frames. Cameras, edge boxes and production lines already produce frames: send them with their timestamps, and the model reads them as one video, in order and with timing, so it can answer what happens over time: how many cars passed, which way a vehicle turned, whether smoke appeared, whether the conveyor stopped. We take frames, not video files or links, and a small helper cuts a local file into frames on your own machine.

When should I use a session?

More than 10 decisions a second? Use a session. It is built for loops: survey, inspection and delivery drones, simulators and games, machine monitoring, long-running agents. You open it once with the instructions, the fixed state and the questions, then stream frames (text, JSON or an image) over one WebSocket and get a typed decision back for each frame. Your context stays loaded for the whole session, so each frame costs only its own tokens, and a client that falls behind gets the newest frame answered, never a backlog, with the skipped frames reported. A session is a decision layer that feeds your controller: the vehicle's own control loop keeps flying it.

Yes. The same request to the same model version returns the same answer, every time. There is no sampling, so your tests hold and every logged decision can be replayed.

Get an API key, paste one curl and branch on the typed answer. The quickstart takes under five minutes. To try a request in your browser first, the playground runs it with a free account.

What data do you keep?

Requests are processed to return your answers and to run the service. Unless you opt out in the console, they also help improve the models, with identifiers removed first; with zero retention, on an Order Form or under a Defence Contract Addendum they never do. Inputs and answers are deleted 30 days after each request by default, and zero retention is available on request on every plan. The Privacy Policy and the Data Processing Addendum have the details; questions go to privacy@bynn.com.

decision model + inference api

Typed decisions in milliseconds.

Our decision model is a System One model: text, JSON or images in, a calibrated choice, score, truth, number, point or box out, in about 5 ms on DecisionNode Flash.

Get API key

Open playground

$0.042 per million input tokens

output is free / no sales call

example

DecisionNode Flash

choice

decisionnode-1.0-flash

...

“Charged twice and no reply yet.”

Where should this ticket go?

billing...
bug...
account...

auto-refund?acts at 0.80

refund...

confidence...

acts at 0.70deciding

Billing, refund sent

billing + refund

Routed to its queue

otherwise

1,284 actions taken

61 re-checked first

example decisionsp50 5 ms

09:41:02sort parcelreturns 0.83
09:41:01check chargefraud 0.12 false4 ms
09:40:58moderate postseverity 0.414 ms

New ticketwebhook

decision model + inference api

Typed decisions in milliseconds.

Our decision model is a System One model: text, JSON or images in, a calibrated choice, score, truth, number, point or box out, in about 5 ms on DecisionNode Flash.

Get API key

Open playground

$0.042 per million input tokens

output is free / no sales call

example

DecisionNode Flash

choice

decisionnode-1.0-flash

...

“Charged twice and no reply yet.”

Where should this ticket go?

billing...
bug...
account...

auto-refund?acts at 0.80

refund...

confidence...

acts at 0.70deciding

Billing, refund sent

billing + refund

Routed to its queue

otherwise

1,284 actions taken

61 re-checked first

example decisionsp50 5 ms

09:41:02sort parcelreturns 0.83
09:41:01check chargefraud 0.12 false4 ms
09:40:58moderate postseverity 0.414 ms

New ticketwebhook

one HTTPS call from anything you already run

any http client / POST /v1/decide

typescriptts/jsdn.decide()
pythonpythondn.decide()
gogodn.Decide()
rubyrubydn.decide
/v1/decideDecisionNode
curlcurl-X POST
mcpmcptools/call
webhookswebhookson event
agentsagentsbefore acting

same order, two engines

Sixteen answers before the first word.

The same order and the same sixteen checks. DecisionNode answers them all in one pass while a chat LLM is still writing.

truthIs this statement true? One calibrated probability from 0.00 to 1.00 that the statement is true. 0.80 means true about 8 times in 10. Your code picks the cut-off: 0.50 for a plain yes or no, higher when acting on a false yes is costly.

DecisionNode Flash

POST /v1/decide, one call

request200 OK16 questions, one callmodel decisionnode-1.0-flash
fraud_risktruthTasked0.04 false
address_validtruthTasked0.97 true
gift_ordertruthTasked0.93 true
ship_lanechoiceCaskedstandard 0.88
carrierchoiceCaskedcourier 0.81
priorityscoreSasked1.8 of 3
fragile_itemstruthTasked0.91 true
needs_signaturetruthTasked0.22 false
customs_formtruthTasked0.03 false
promo_abusetruthTasked0.71 true
duplicate_ordertruthTasked0.06 false
stock_reservedtruthTasked0.99 true
packagingchoiceCaskedpadded 0.76
delivery_riskscoreSasked0.6 of 4
split_shipmenttruthTasked0.12 false
next_actionchoiceCaskedship 0.84
usagepending410 tokens in, 0 out

a chat LLM

chat completion, "reply in JSON"

time to all 16 answers: 25 ms against 3.78 s
cost of this order: 0.0014¢ against 0.01¢
price ratio: about 8.8× cheaper

DecisionNode Flash against GPT-5.6 Luna, the chat LLM timed on our benchmarks page. Time: DecisionNode Flash's 19 ms end-to-end median plus 6 ms for the extra questions; GPT-5.6 Luna's first token at 630 ms, then 126 output tokens at 25 ms each. Cost is tokens times list price: 410 input tokens at $0.035 per 1M, output free, against 756 input tokens at $0.1 and 126 output tokens at $0.4 per 1M. Mock timings. Preliminary results, October 2026.

See every model compared

Race: sixteen checks on one order, DecisionNode Flash against a chat LLM
check	DecisionNode Flash, typed answer	a chat LLM, text answer
Is this order fraud? (truth)	0.04 false	"fraud_risk": "low"
Is the address deliverable? (truth)	0.97 true	"address_valid": "yes"
Is this a gift? (truth)	0.93 true	"gift_order": "true"
Which ship lane? (choice)	standard 0.88	"ship_lane": "standard shipping"
Which carrier? (choice)	courier 0.81	"carrier": "courier"
How urgent, 0 to 3? (score)	1.8 of 3	"priority": "medium-high"
Any fragile items? (truth)	0.91 true	"fragile_items": "yes, the lamp"
Needs a signature? (truth)	0.22 false	"needs_signature": "probably not"
Needs a customs form? (truth)	0.03 false	"customs_form": "no"
Is the promo code abused? (truth)	0.71 true	"promo_abuse": "possibly"
Is this a duplicate order? (truth)	0.06 false	"duplicate_order": "no"
Is the stock reserved? (truth)	0.99 true	"stock_reserved": "yes"
Which packaging? (choice)	padded 0.76	"packaging": "padded box"
Delivery risk, 0 to 4? (score)	0.6 of 4	"delivery_risk": "low"
Split into two parcels? (truth)	0.12 false	"split_shipment": "false"
Ship, hold, route or flag? (choice)	ship 0.84	"next_action": "ship it"

time to all 16 answers: 25 ms against 3.78 s
cost of this order: 0.0014¢ against 0.01¢
price ratio: about 8.8× cheaper

the decision model

A model built to decide.

Hand-written rules break on the first message nobody planned for. Chat models word the same answer differently every time. DecisionNode returns a typed answer and how sure it is, the same way every time.

How a decision runs

in
State and questions
text, JSON or images
call
One request
POST /v1/decide
model
Our model on our GPUs
one pass, no text generated
out
Typed answers
choice, score, truth, number, point or box
branch
Your code branches
confidence < 0.70

Six ways to ask, one typed answer each

Every type names its answer: choice (one of your labels), score (a level on your scale), truth (a probability), number (a value on your grid), point (a spot in your image), box (a rectangle around the thing). Explore each type below:

state“Customer: I was charged twice and nobody has replied for 3 days.”

question

"route": {
  "type": "choice",
  "criteria": {
    "billing": "money",
    "bug": "broken",
    "account": "login"
  }
}

answer

"route": {
  "type": "choice",
  "choice": "billing",
  "confidence": 0.81,
  "probabilities": {
    "account": 0.05,
    "billing": 0.87,
    "bug": 0.08
  }
}

question

"urgency": {
  "type": "score",
  "criteria": [
    "routine", "today",
    "urgent", "critical"
  ]
}

answer

"urgency": {
  "type": "score",
  "score": 2.31,
  "confidence": 0.47,
  "probabilities": {
    "0": 0.01, "1": 0.09,
    "2": 0.48, "3": 0.42
  }
}

question

"refund": {
  "type": "truth",
  "instructions":
    "Refund this automatically?",
  "criteria": {
    "true":
      "Charged twice for the same order",
    "false":
      "Any other billing issue"
  }
}

answer

"refund": {
  "type": "truth",
  "truth": 0.94
}

question

"days_waiting": {
  "type": "number",
  "instructions":
    "Days without a reply?",
  "min": 0,
  "max": 30
}

answer

"days_waiting": {
  "type": "number",
  "number": 3,
  "expected": 3.06,
  "confidence": 0.88,
  "probabilities": {
    "2": 0.04, "3": 0.88,
    "4": 0.06, "5": 0.02
  }
}

question

"pick": {
  "type": "point",
  "instructions": "Point to the yellow package closest to the robot arm.",
  "image": "belt"
}

answer

"pick": {
  "type": "point",
  "x": 812,
  "y": 455,
  "present": 0.97,
  "confidence": 0.91,
  "image": "belt",
  "width": 1920,
  "height": 1080
}

question

"damage": {
  "type": "box",
  "instructions": "Draw a box around the damaged package.",
  "image": "belt"
}

answer

"damage": {
  "type": "box",
  "x1": 300,
  "y1": 120,
  "x2": 520,
  "y2": 410,
  "present": 0.95,
  "confidence": 0.88,
  "image": "belt",
  "width": 1920,
  "height": 1080
}

returns choice, probabilities, confidenceTry it in the playground

One request answers every question

Send the state once with every question you have. It is read in one pass and shared, so ten questions cost about the same as one.

POST /v1/decide

stateread once0.0 ms62 tokens
routechoicebilling
urgencyscore2.31 urgent
refundtruth0.94 true

zoomin parallelroute+0.4 msurgency+0.5 msrefund+0.3 ms

62 input tokens, 0 output1 pass, no text generated

Confidence you can act on

Every answer says how sure the model is, and the numbers are calibrated. Set a line once: above it your code acts instantly. Below it, the full model checks the same request again, or your code takes the safe path. Nobody waits.

ticket 4471routeact at

DecisionNode Flashbilling, 4.8 msbelow line
re-checkDecisionNode MoEbilling, +8.1 msabove line
actedmoved to billing, reply sent12.9 ms

>if (flash.confidence >= 0.70) act()>else if (full.confidence >= 0.70) act()>else safePath()

Text and images in one call

Attach up to 16 photos to the same request: a receipt, a camera frame, a parcel on the belt. The model reads printed text, amounts, objects and layout, and answers with the same question types. Send video as frames with their timestamps, and the model reads them as one clip.

Photo of a till receipt from Kaffebar Norr in Stockholm, total 188,00 kr with 12% moms — which currency is on this receipt?choice, image + text

which currency is on this receipt?choice, image + text

point and box

Ask where, get coordinates

Two new question types answer where. A point is one spot in the image, a box is the rectangle around the thing: coordinates your code can draw, crop or hand to a machine, each with how likely the thing is there and how likely the place is right.

what you could ask

Production lines and sorting"point"“Point to the next yellow package to pick.”
Quality control"box"“Draw a box around the scratch on the panel.”
Surveillance"box"“Draw a box around the person climbing the fence.”
Traffic"box"“Draw a box around the car that stopped in the crossing.”
Warehouses"point"“Point to the empty slot on the second shelf.”
Drones and robots"point"“Point to the landing marker.”

From pixels to motion

belt camera, line 3 · 1920 × 1080

pick

damage

01Camera frame
02Point + box
03Your controller
04Robot action

The belt advances. One camera frame carries two questions.

Illustrated playback · pause to inspect each handoff

JSON response

"pick": {
  "type": "point",
  "x": 812,
  "y": 455,
  "present": 0.97,
  "confidence": 0.91
}

"damage": {
  "type": "box",
  "x1": 300,
  "y1": 120,
  "x2": 520,
  "y2": 410,
  "present": 0.95,
  "confidence": 0.88
}

[ a model company and an inference company ][ model company, inference company ]

We build the decision modeland the machine it runs on.

one tick, one decision

now

inference

Fast because we own the path.

Our own decision model on our own inference stack and GPUs. No third-party API between your request and the answer.

Every hop is ours

One pass, no text generation, nothing in between. DecisionNode Flash answers a short request in about 5 ms.

the path of one request

requestyour app
read oncetext, images
questionsat once
decideour stack

no sampling: same answer every time

One fleet, two models

Both models run on our own GPUs behind one API, priced per 1M input tokens. Switch with the model field; nothing else changes.

an example traffic mix

DecisionNode MoE
decisionnode-1.0
p50
9 ms
per 1M in
$0.042
share
61%
DecisionNode Flash
decisionnode-1.0-flash
p50
5 ms
per 1M in
$0.035
share
39%

same request, same answer

Done before an LLM writes a word

A chat LLM writes its answer token by token and needs half a second to four seconds. DecisionNode reads the input once and answers in one pass, fast enough for every request, every payment, every camera frame.

End to end from one client, network included, one short decision; the chat LLM on the same decision. Preliminary results, October 2026.

time to a typed answer, end to end, log scale

DecisionNode Flash
5 ms on our side
DecisionNode MoE
9 ms on our side
a chat LLM
chat LLM 1,300 ms

Time to a short typed answer, end to end from one client, on a log scale: DecisionNode Flash 19 ms, DecisionNode MoE 24 ms, a chat LLM 1,300 ms.
model	time to answer	note
DecisionNode Flash	19 ms	5 ms on our side
DecisionNode MoE	24 ms	9 ms on our side
a chat LLM	1,300 ms	a chat LLM, 1,300 ms

up to 68× faster on DecisionNode Flash

Paste your own request and watch the clock.

How we measured Open playground

sessions

A decision for every frame.

Open one session, send the fixed part once, then stream frames over a WebSocket. Every frame comes back as a typed decision, a decision layer that feeds your controller.

POST /v1/sessionssession opening

sent once instructions, mission brief, geofence, 2 questions, window 8, model decisionnode-1.0

each frame bills only its own tokens

Survey drone, line inspection

frames in0

/v1/sessions/ses_7f3a/stream

frametelemetry JSON; text and images work too

window 8leaves context

contextsent once

Typed decisions

decisions out0

waiting for the first frame

More than 10 decisions a second? Use a session.

Same models, same typed and calibrated answers, the same safety check on every frame. The context is billed once when the session opens; each frame, its own tokens.

Read the sessions docs

A survey drone streams telemetry frames into one session; every frame is answered with a flight mode and an abort probability
frame	telemetry	mode: Which flight mode now?	abort: Should the mission be aborted now?
1	battery 0.64, wind 4.2 m/s, 155 m from home	continue 0.93, confidence 0.91 (continue 0.93, hold 0.04, return 0.02, land 0.01)	0.02 false
2	battery 0.63, wind 4.6 m/s, 225 m from home	continue 0.93, confidence 0.91 (continue 0.93, hold 0.04, return 0.02, land 0.01)	0.02 false
3	battery 0.61, wind 5.1 m/s, 295 m from home	continue 0.92, confidence 0.89 (continue 0.92, hold 0.05, return 0.02, land 0.01)	0.02 false
4	battery 0.60, wind 5.8 m/s, 365 m from home	continue 0.91, confidence 0.88 (continue 0.91, hold 0.05, return 0.03, land 0.01)	0.03 false
5	battery 0.58, wind 6.6 m/s, 435 m from home	continue 0.88, confidence 0.84 (continue 0.88, hold 0.07, return 0.04, land 0.01)	0.03 false
6	battery 0.56, wind 8.9 m/s, 500 m from home	continue 0.71, confidence 0.61 (continue 0.71, hold 0.22, return 0.06, land 0.01)	0.06 false
7	battery 0.54, wind 11.8 m/s, 530 m from home	hold 0.62, confidence 0.49 (continue 0.27, hold 0.62, return 0.09, land 0.02)	0.14 false
8	battery 0.52, wind 12.6 m/s, 535 m from home	hold 0.71, confidence 0.61 (continue 0.19, hold 0.71, return 0.08, land 0.02)	0.17 false
9	battery 0.51, wind 12.2 m/s, 535 m from home	latest wins: skipped, a newer frame came in
10	battery 0.50, wind 11.4 m/s, 540 m from home	latest wins: skipped, a newer frame came in
11	battery 0.48, wind 9.6 m/s, 565 m from home	continue 0.58, confidence 0.44 (continue 0.58, hold 0.33, return 0.07, land 0.02)	0.09 false
12	battery 0.45, wind 9.2 m/s, 635 m from home	continue 0.69, confidence 0.59 (continue 0.69, hold 0.23, return 0.06, land 0.02)	0.06 false
13	battery 0.42, wind 10.1 m/s, 705 m from home	continue 0.62, confidence 0.49 (continue 0.62, hold 0.21, return 0.15, land 0.02)	0.06 false
14	battery 0.39, wind 10.6 m/s, 770 m from home	continue 0.44, confidence 0.25 (continue 0.44, hold 0.17, return 0.37, land 0.02)	0.07 false
15	battery 0.36, wind 11.0 m/s, 835 m from home	return 0.52, confidence 0.36 (continue 0.31, hold 0.14, return 0.52, land 0.03)	0.08 false
16	battery 0.34, wind 11.2 m/s, 830 m from home	return 0.71, confidence 0.61 (continue 0.17, hold 0.09, return 0.71, land 0.03)	0.08 false
17	battery 0.31, wind 11.4 m/s, 820 m from home	return 0.84, confidence 0.79 (continue 0.09, hold 0.05, return 0.84, land 0.02)	0.07 false
18	battery 0.29, wind 11.9 m/s, 750 m from home	return 0.88, confidence 0.84 (continue 0.06, hold 0.04, return 0.88, land 0.02)	0.06 false
19	battery 0.27, wind 12.3 m/s, 670 m from home	return 0.90, confidence 0.87 (continue 0.05, hold 0.03, return 0.90, land 0.02)	0.06 false
20	battery 0.25, wind 12.8 m/s, 585 m from home	return 0.91, confidence 0.88 (continue 0.04, hold 0.03, return 0.91, land 0.02)	0.06 false

boosters

Specialist models read your pictures first.

Name one beside your questions. It reads every picture before the decision, and DecisionNode weighs its reading as evidence in the same request. Priced per picture read.

A seller's selfie: a smiling woman with long dark hair, in a living room full of plants.

Human Ageage
$0.024 a picture
Age-gate a store from a selfie, and ask for an ID only when the answer is close.
Estimates a person's age from a face in the picture, using the agemin.com (opens in a new tab) age estimation model, ranked #1 by NIST for Child Online Safety (ages 13 to 16) on true positive rate.
reads a picture of a person; returns age_years, low, high, confidence, faces
AI-generatedai_generated
$0.006 a picture
Catch a generated profile photo before it goes live.
Whether a picture was made by an image generator, and which kind, using Trinity, the state-of-the-art AI image detection model from detector24.ai (opens in a new tab).
reads any picture; returns generated, confidence, method, top_generator, c2pa_signer, c2pa
Face detectionface_detection
$0.001 a picture
Count the faces in a frame and see where each one is.
Finds the faces in a picture: how many it shows and where each one is, with a box and a confidence per face, from Bynn's face detection model.
reads any picture; returns faces, detections
Face recognitionface_recognition
from $0.001 a picture
Recognise the faces you enroll once, your staff or your members, or a public figure on stage.
Matches the faces in a picture against a collection: one of your own, enrolled through the API or the console, or the public celebrities collection. Each match carries a score and the item's reference, with your label or the public figure's name, and never a picture.
reads a picture of a person; returns faces, collection_name, matches
OCRocr
$0.0015 a picture
Read every word of an invoice before deciding on it: plain text, a table, a formula or the fields of your JSON schema.
The text a picture shows, read by Bynn OCR: all of it by default, or what a prompt asks for, such as a formula in LaTeX, a table in HTML or Markdown, or the fields of a JSON schema. Full support for English and Chinese; other languages need a prompt that names them.
reads any picture; returns text, truncated
Depth Visionbynn_depth
free
How far away things are, in meters: straight ahead, in seven directions and cell by cell.
Distance in meters to what's in the picture: the nearest obstacle straight ahead, seven directions from left to right, and a grid of up to 16 by 16 cells.
reads any picture; returns ahead, directions, ground, grid_rows, grid_cols, grid, depth_min, depth_max, depth_median, width, height
Segment Visionbynn_segment
$0.001 a picture
Where the things you name are: each one's outline, box and centre, and with depth on, how far away and how large it is.
Finds what you name in words in the picture: each match's outline, box and centre, and with depth on, its distance and size in meters.
reads any picture; returns prompts, instances, depth_source, assumed_hfov_deg, width, height

Name a booster beside the questions it informs; its reading comes back with the answer.

How boosters work Run the seller check

benchmarks

How accurate it is, measured.

15 suites of real decisions, from policy rules to web tasks and prompt injection. Every system answered the same items.

15-suite mean

Mean accuracy over every text suite, one dot per system.

25,134 items

Preliminary results, October 2026

3Cloudflare Clef 27B0.786
5Cloudflare Clef-flash 9B0.759
7Cygnet (Gemma 4 12B)0.717
1DecisionNode MoEours0.816
4DecisionNode Flashours0.784
2TypeSafe Jev 1.13^{, note 1}0.803
8Kev-4B0.717
6pplx-decider-27b0.746
9Strands Decider 2B^{, note 2}0.622

accuracy; on a suite, the line across a dot is its 95% interval

Note 1, TypeSafe Jev 1.13: called through its official API. Reads text only, so it has no image results.
Note 2, Strands Decider 2B: its published training data contains items of two of our test sets.

where we lead

Pick a suite to rank every system.

where we trail

DecisionNode MoEJevline length is the gap in points

DecisionNode Flash, end to end

19 ms

5 ms of it on our side

up to 68× faster than GPT-5.6 Luna, a chat LLM, on the same decision, timed end to end from one client.

15 text suites, 25,134 items, development splits. 95% intervals for our models and Jev from a bootstrap that resamples whole inputs. Jev was called through its official API; open models ran on our GPUs as their authors serve them. October 2026.

DecisionNode is not affiliated with TypeSafe AI, Cloudflare, OpenAI or the other model publishers. Their names identify the systems measured.

See every suite

pricing

You pay for input, not for output.

Answers generate no text, so output costs nothing. Pay per million input tokens from a prepaid balance. No seats, no minimums.

estimate at list prices

How many input tokens do you send a month?

about 2,000,000 requests at 500 input tokens each

DecisionNode MoE, a month

DecisionNode Flash, a month

output tokens$0.00

batch jobs30% to 50% off

Get API key

DecisionNode Flash
$0.035
per 1M input tokens
batch jobs $0.0175, 50% off
decisionnode-1.0-flash
- our fastest tier
- a lower price per token
- the same answer shapes
- one field to switch models
Get API key
DecisionNode MoE
full model
$0.042
per 1M input tokens
batch jobs $0.0294, 30% off
decisionnode-1.0
- our most accurate model
- best on subtle calls
- images and video at $0.09 per 1M tokens
- the playground default
Get API key
Dedicated
Custom
capacity reserved for you
both models
- capacity on our own GPUs
- custom rate limits
- same API, no code changes
- a direct line to our engineers
Talk to us

both models

text and up to 16 images per request
up to 64k tokens of context
output always free
one API

images and video

$0.09 per 1M image and video tokens on DecisionNode MoE, $0.08 on DecisionNode Flash
no per-image fee
about one token per 32 × 32 pixels: a 1,024 × 1,024 image is about 1,026 tokens
a video frame at the default size is about 64 tokens

compared

OpenAI Decisions API: $0.10 per 1M input tokens, images $0.10 per 1M
DecisionNode MoE at 42% of that text price, DecisionNode Flash at 35%

free tier

50 DecisionNode Flash requests, 20 DecisionNode MoE requests and 10 booster calls a day on your first workspace, playground and API keys alike, renewed at 00:00 UTC

faq

Questions developers ask.

The short answers. The docs have the long ones, with the request and response for every case.

Read the docs

A chat model writes its answer token by token, so a short answer takes half a second to several seconds and rarely answers the same way twice. DecisionNode reads your input once and answers your typed questions in one pass, each answer with how sure the model is: a choice, a score, a truth or a number. It generates no text, so it answers in milliseconds. A System One model is a class of AI model that returns typed, probabilistic decisions from a fixed answer space instead of generating text. Given a piece of state and a set of questions defined in code, it answers each one with a schema-valid value such as a chosen option, a score, or a probability, and then stops. That is the right tool for routing, triage, moderation, scoring and gating, and the wrong one for writing.
Every type names its answer: choice (one of your labels), score (a level on your scale), truth (a probability), number (a value on your grid), point (a spot in your image), box (a rectangle around the thing).
Choice picks one option from the list you send and returns a probability for each.
Score places the input on your ordered scale, for example routine to critical, and returns the expected level.
Truth, truth in the API, asks whether a statement is true: one calibrated probability from 0.00 to 1.00, so 0.80 means true about 8 times in 10. Your code picks the cut-off: 0.50 for a plain yes or no, higher when acting on a false yes is costly.
Number, number in the API, asks: How many, or what value? One value on a grid you set (min, max and an optional step), with a calibrated probability for every value: the most probable value, the expected value and the confidence. A count of cars in a photo, overdue invoices in a statement, a year, a rating out of ten.
Point, point in the API, and box, box, answer where. A point is one spot in the image, as x and y, with the probability that the thing is there at all (present) and that the spot is on it (confidence). A box is a rectangle around the thing, as its top-left corner (x1, y1) and bottom-right corner (x2, y2), with present and confidence.
You pay per million input tokens from a prepaid balance: $0.042 on DecisionNode MoE, $0.035 on DecisionNode Flash. Image and video tokens are priced per million too, $0.09 on DecisionNode MoE and $0.08 on DecisionNode Flash, with no fee per image. Work that can wait runs as a batch, 30% off the live price on DecisionNode MoE and 50% off on DecisionNode Flash. The model writes no text, so every response reports output_tokens: 0 and output never costs anything. No seats, no minimums.
A repeated state is billed in full each time, with no discount for input sent before. Your first workspace has a free tier: 50 DecisionNode Flash requests, 20 DecisionNode MoE requests and 10 booster calls every day, from the playground and from API keys alike.
Yes. Two question types, point and box, answer where. Ask "Point to the yellow package closest to the robot arm" or "Draw a box around the person at the door" and get coordinates you can draw, crop or hand to a machine. Every answer also says how likely it is that the thing is there at all (present), so your code acts on a sure answer and takes the safe path on an unsure one. Each place carries a confidence too, not calibrated yet, so act on present. The instructions can carry conditions and refer to the state and to other images, which a fixed list of detector classes cannot: production lines and sorting, quality control, surveillance and more.
Billed by the rows it runs, like any question: the input tokens of its presence row and of the coordinate rows of the place it answers (two for a point, four for a box), about 140 tokens for a point and 250 for a box, plus the image once per request at the image and video rate. No fee per point or box, and output is free.
- Read about point and box
Yes, from camera images and from video sent as frames. Monitoring areas comes first: traffic and its flow (counts, turns, congestion, wrong-way drivers, pedestrians crossing), people in an area (counts, flow, crowding) and incidents (break-ins, theft, fire and smoke). Then production lines and sorting: one decision per item at the sorter (damaged, mislabelled, upside down, which lane), counting products, spotting jams and stoppages. Drones can be monitored from above through their camera frames, and a session hands a drone's own controller a typed decision for every frame. Warehouse machines get command-level decisions the same way: go, stop, which aisle.
Need it on your own hardware, on premises, at the edge or next to the cameras? For qualifying deployments, talk to us at hello@bynn.com.
- Industry templates
- Talk to us
Yes, sent as frames. Cameras, edge boxes and production lines already produce frames: send them with their timestamps, and the model reads them as one video, in order and with timing, so it can answer what happens over time: how many cars passed, which way a vehicle turned, whether smoke appeared, whether the conveyor stopped. We take frames, not video files or links, and a small helper cuts a local file into frames on your own machine.
- Read about video
More than 10 decisions a second? Use a session. It is built for loops: survey, inspection and delivery drones, simulators and games, machine monitoring, long-running agents. You open it once with the instructions, the fixed state and the questions, then stream frames (text, JSON or an image) over one WebSocket and get a typed decision back for each frame. Your context stays loaded for the whole session, so each frame costs only its own tokens, and a client that falls behind gets the newest frame answered, never a backlog, with the skipped frames reported. A session is a decision layer that feeds your controller: the vehicle's own control loop keeps flying it.
- Read the sessions docs
Get an API key, paste one curl and branch on the typed answer. The quickstart takes under five minutes. To try a request in your browser first, the playground runs it with a free account.
Requests are processed to return your answers and to run the service. Unless you opt out in the console, they also help improve the models, with identifiers removed first; with zero retention, on an Order Form or under a Defence Contract Addendum they never do. Inputs and answers are deleted 30 days after each request by default, and zero retention is available on request on every plan. The Privacy Policy and the Data Processing Addendum have the details; questions go to privacy@bynn.com.
- Privacy Policy
- Data processing

Typed decisions in milliseconds.

Typed decisions in milliseconds.

How a decision runs

Six ways to ask, one typed answer each

One request answers every question

Confidence you can act on

Text and images in one call

Ask where, get coordinates

what you could ask

We build the decision modeland the machine it runs on.

Every hop is ours

One fleet, two models

Done before an LLM writes a word

Human Ageage

AI-generatedai_generated

Face detectionface_detection

Face recognitionface_recognition

OCRocr

Depth Visionbynn_depth

Segment Visionbynn_segment

15-suite mean

where we lead

where we trail

DecisionNode Flash

DecisionNode MoE

Dedicated

images and video

compared

free tier

Questions developers ask.

What is a decision model, and how is it different from a chat model?

What are the question types?

How does pricing work, and why is output free?

Which model should I pick, DecisionNode MoE or DecisionNode Flash?

Can it read images?

Can it say where something is in an image?

Can it watch traffic, people or a production line?

Can it watch video?

When should I use a session?

Is it deterministic?

How do I start?

What data do you keep?

How a decision runs

Six ways to ask, one typed answer each

One request answers every question

Confidence you can act on

Text and images in one call

Ask where, get coordinates

what you could ask

We build the decision modeland the machine it runs on.

Every hop is ours

One fleet, two models

Done before an LLM writes a word

Human Ageage

AI-generatedai_generated

Face detectionface_detection

Face recognitionface_recognition

OCRocr

Depth Visionbynn_depth

Segment Visionbynn_segment

15-suite mean

where we lead

where we trail

DecisionNode Flash

DecisionNode MoE

Dedicated

images and video

compared

free tier

Questions developers ask.

What is a decision model, and how is it different from a chat model?

What are the question types?

How does pricing work, and why is output free?

Which model should I pick, DecisionNode MoE or DecisionNode Flash?

Can it read images?

Can it say where something is in an image?

Can it watch traffic, people or a production line?

Can it watch video?

When should I use a session?

Is it deterministic?

Human Age`age`

AI-generated`ai_generated`

Face detection`face_detection`

Face recognition`face_recognition`

OCR`ocr`

Depth Vision`bynn_depth`

Segment Vision`bynn_segment`

Human Age`age`

AI-generated`ai_generated`

Face detection`face_detection`

Face recognition`face_recognition`

OCR`ocr`

Depth Vision`bynn_depth`

Segment Vision`bynn_segment`