| Plan | Fee | Usage included | Includes |
|---|---|---|---|
| Starter | $5 a month or $50 a year | $5 a month, up to about 280,000 short text decisions* | OCR, face and AI‑image detection in production · email support · 3 keys, 2 members |
Pro recommended | $29 a month or $290 a year | $35 a month, up to about 2 million short text decisions* | Batch jobs and sessions · knowledge bases and face collections · 5 more boosters in production · 10 keys, 5 members |
| Business | $99 a month or $990 a year | $120 a month, up to about 6.8 million short text decisions* | Everything in Pro · fine-tuning · priority support · 100 keys, 25 members |
| Dedicated | Reserved capacity on our own GPUs, custom rate limits and negotiated rates, invoiced monthly: talk to us | ||
* A short text decision is one request of 500 input tokens on DecisionNode Flash at its list rate, with no boosters. Longer inputs, images and DecisionNode buy fewer; the estimator on the pricing page works out your own.
Plans#
Plans are paid monthly or yearly, where a year costs ten months (two months free). A plan's fee is a monthly usage allowance at the per-token rates on this page: the tokens, images, video and booster calls your requests use draw it first, and past it, credit you bought. The rates are the same on every plan; the plan sets the allowance, the rate ceilings, the number of keys and members, and the features below. You choose and change a plan on the console's Plan page.
| (unavailable)-day trial | Starter | Pro | Business | |
|---|---|---|---|---|
| Usage included | (unavailable) for the trial | $5 a month | $35 a month | $120 a month |
| Booster test calls, boosters outside the plan | (unavailable) a day | (unavailable) a day | (unavailable) a day | (unavailable) a day |
| Boosters in production | (unavailable)'s | A first set | Most, Knowledge included | Every booster |
| Batch jobs and sessions | Yes | No | Yes | Yes |
| Knowledge bases and face collections | Yes | Knowledge for testing | Yes | Yes |
| Fine-tuning | No | No | No | Yes |
| Requests a second | (unavailable)'s | 5 | 25 | 100 |
| API keys and members | (unavailable)'s | 3 keys, 2 members | 10 keys, 5 members | 100 keys, 25 members |
The (unavailable)-day (unavailable) trial#
- Starts at sign-up, card required. A new workspace's Owner adds a card at Stripe Checkout and gets (unavailable), with every (unavailable) feature and (unavailable)'s rate limits, for (unavailable) days. Nothing is charged during the trial.
- A (unavailable) usage cap. When it is used, requests answer
402with the reasontrial_allowance_useduntil the trial ends, or until Start (unavailable) now in the console ends the trial at once and charges (unavailable)'s first month with its full allowance. - Converts to (unavailable). At the end of the (unavailable) days the trial becomes (unavailable) at (unavailable) a month, unless you switch plan or cancel first. From day one the console shows the days left, the usage of the cap, and the date and amount of the first charge; a notice and an email come three days and one day before it.
- Switch or cancel any time before. On the Plan page, Switch to Starter keeps the trial and converts it to Starter instead, and Cancel trial lets it run to its last day and charges nothing. The workspace then has no plan until one is chosen.
- One trial per card and per person. A card or a person that has had a trial starts (unavailable) at once, charged, with the notice "This card has had a trial; your (unavailable) plan starts today".
- No trial needed. The same step starts Starter or Business at once, monthly or yearly.
The allowance#
- Granted each period. A monthly plan gets its allowance at each renewal; a yearly plan gets it every month of its year, on the day the plan started.
- Drawn first. Usage draws the allowance before any credit you bought, so purchased credit is only spent once the month's allowance is used.
- Does not roll over. What is left of the allowance when the period ends lapses: it is not carried over or refunded. The console's Plan page shows what is used and the day it resets.
- Notices. The console tells the workspace when 80%, 95% and all of the allowance are used, with the plan that would cover the month and its fee, the reset date and what happens next.
Past the allowance#
Usage past the allowance draws purchased credit, which lasts twelve months from its purchase. Auto reload is on by default on a plan: it buys credit when the balance runs low, under a monthly cap equal to the plan's monthly fee, which you can raise, remove, or turn off with auto reload under Billing. When the allowance and the credit are both used, requests answer 402 with the reason allowance_used until credit is added, the cap is raised, the plan is upgraded, or the period resets.
Test calls#
A booster your plan does not run in production still runs through the daily test calls, (unavailable) a day on every plan and on the trial, which reset at 00:00 UTC and cost nothing. Once they are used, it answers 403 plan_required, naming the plan that includes it. Boosters your plan includes are billed from the first call.
What is free#
Boosters outside your plan stay free to test through the test calls. A workspace has no other free daily counts.
Change or cancel#
- Upgrades take effect at once. You pay the new plan's fee less a credit for the allowance you have not used this period, valued at your current fee's rate: with all of Starter's $5 used, Pro costs its full $29; with a fifth of it used, four fifths of Starter's $5 comes off. On a yearly plan the whole months left in the year count too. The credit is never more than the new fee, purchased credit is never part of it, and the console shows the amount before you pay. You get the new plan's full allowance for a new period, and what was left of the old one lapses.
- Downgrades take effect at the period end. You keep the current plan and its allowance until then; the Plan page shows the change waiting, and Keep my plan undoes it.
- Cancelling also takes effect at the period end. The fee for the period is not refunded, and credit you bought stays in the workspace.
- When a plan ends, open sessions close, running batch jobs finish, and deployed fine-tuned models are undeployed after seven days' notice and kept in storage. A workspace with no plan answers every decision with
402 plan_required, and its console opens on the start step until a plan is chosen. - Plans are changed in the console only, by an Owner, Admin or Billing member. A Developer asks for an upgrade from the console, and the request reaches them.
Per-token rates#
The same rates on every plan and on the trial: the allowance buys them, and past it, so does credit you buy.
| Model | Family id you send as model | Text / 1M | Images and video / 1M | Batch text / 1M |
|---|---|---|---|---|
| DecisionNode | decisionnode | $0.042 | $0.09 | $0.0294 |
| DecisionNode Flash | decisionnode-flash | $0.035 | $0.08 | $0.0175 |
| Dedicated | Reserved capacity on our own GPUs, custom limits: talk to us | |||
Prices are in USD. A price change applies to usage from its date; within 30 days of that date it shows here beside the price, "from" the date. Batch jobs are 30% off the live price on DecisionNode and 50% off on DecisionNode Flash, on text, image and video tokens alike. Images and video have no fee per image: you pay for their tokens only.
DecisionNode Edge, the open-weight System One models you run on your own hardware, has no price: the weights are free to download under AGPL-3.0, with a commercial licence for closed products and services. Only the questions it sends on to the hosted API are billed, at the prices above.
Images and video#
Image and video tokens are billed at the image and video price per million, $0.09 on DecisionNode and $0.08 on DecisionNode Flash, with no fee per image; the state and the questions of the same request are billed at the text price. An image costs about one token per 32 x 32 pixels after the server has scaled it, so the size you send sets the cost.
| Input | Tokens | On DecisionNode | On DecisionNode Flash |
|---|---|---|---|
| A 1,024 x 1,024 image | About 1,026 | About $0.0000923 | About $0.0000821 |
| A video frame at the default size | About 64 | About $0.0000058 | About $0.0000051 |
usage.input_tokens counts image and video tokens together with the text, and the x-decisionnode-image-tokens response header says how many came from images.
Compared with OpenAI's Decisions API#
| API | Text / 1M | Images / 1M | Text price against OpenAI's |
|---|---|---|---|
| DecisionNode | $0.042 | $0.09 | 42% |
| DecisionNode Flash | $0.035 | $0.08 | 35% |
| OpenAI Decisions API | $0.10 | $0.10 | 100% |
OpenAI's prices are its public list prices for its Decisions API.
How a request is billed#
cost = usage.input_tokens
× price per token
+ sum of billed_calls × booster price
- usage.input_tokens
- the state, images and every question, as returned in the response
- price per token
- the model's price per million, divided by 1,000,000
- billed_calls × booster price
- for each booster the request named,
evidence.<name>.billed_callstimes that booster's price; none on a request without boosters
usage.output_tokens is always 0. There are no seats, no minimums and no charge per question: the price depends only on what you send. Only answered work is billed: a 200 on /v1/decide, a session opened (201, its fixed part once) and each frame it answers, and the completed lines of a batch. Every error costs nothing: 400, 401, 402, 404, 405, 408, 413, 422, 429, a 500 on our side and 529, so you pay only for requests that return answers. A request refused by the safety check (403) is not charged either. One exception: a request whose boosters read and then failed with a 529 pays for those booster calls; see Boosters.
What each answer says it cost#
Every answer carries its cost in usage.cost: usd, the amount booked for it at your workspace's rates, split into model_usd (its tokens) and boosters_usd (its booster calls). The amounts are numbers in USD with up to 9 decimals. free_tier is false: workspaces have no free daily counts, and a test call adds 0 to boosters_usd. A refused or unbilled answer carries every amount 0. A key with a daily spend cap is never charged more for an answer than was reserved for it: an answer that would cost more is cut to the reservation, so its usage.cost is less than its tokens times the price (Keys on devices). The amounts below are at list prices.
{
"usage": {
"input_tokens": 1121,
"output_tokens": 0,
"booster_tokens": 46,
"boosters": 2,
"cost": {
"usd": 0.03009633,
"model_usd": 0.00009633,
"boosters_usd": 0.03,
"free_tier": false
}
}
}An exact repeat of a request by the same API key within 120 seconds is answered from a cache, in milliseconds, and billed in full like the original (see Repeated requests). A state sent again with new questions is billed at the full price too: there is no discount for input sent before.
A request is billed at the price of the model that answered. When DecisionNode has no ready capacity and DecisionNode Flash answers in its place (the fallback), the request is billed at DecisionNode Flash's price, and the response's model and x-decisionnode-fallback header say so.
Batch jobs#
Work that can wait runs as a batch job at 30% off the live price per token on DecisionNode ($0.0294 per million text tokens) and 50% off on DecisionNode Flash ($0.0175), image and video tokens included. Output stays free, and every line gets the answer the live call would give. Batch jobs take every question type, number questions included.
batch = sum(usage.input_tokens of completed lines)
× batch price per token
- completed lines
- result lines with status 200
- batch price per token
- the live price of each line's model times its batch share (70% on DecisionNode, 50% on DecisionNode Flash)
- Failed, expired and cancelled lines cost nothing for their tokens, and a line the safety check refuses is not charged, as on a live call. Booster calls on a line are charged at the workspace's booster price, never at the batch share, and a line that is not a
200but whose boosters read (the400of an input mismatch, or the529of a required booster) still pays for those booster calls. - Batch results come with a multi-hour delay. There is no set time at which a batch runs or finishes. Requests still unfinished 24 hours after a batch is finalized expire: they are not billed, and you can resubmit them.
- Poll the batch for its state and download the results when it has ended. Nothing is pushed to you.
- Lines draw down the same prepaid balance as live calls. The batch object's
usageshows what a batch has billed so far. - Use batches for backfills, re-scoring after a policy change and re-checks of past decisions. Anything your software is waiting on belongs in a live request.
Boosters#
A booster is priced per call, by booster: age (Human Age) $0.024, ai_generated (AI-generated) $0.006, face_detection (Face detection) $0.001, face_recognition (Face recognition) $0.001 on your own collection and $0.001 on celebrities, ocr (OCR) $0.0015, bynn_depth (Depth Vision) free, bynn_segment (Segment Vision) $0.001 and knowledge (Knowledge) $0.001 a request, drawn from the balance with the request's tokens. A call is one booster reading one picture, so age on three pictures is three calls, $0.072. Each evidence entry says how many calls of its booster were billed (billed_calls). The readings' tokens are input tokens at the model's price, reported in usage.booster_tokens.
| Booster | Reads | Per picture read |
|---|---|---|
age (Human Age) | A picture of a person | $0.024 |
ai_generated (AI-generated) | Any picture | $0.006 |
face_detection (Face detection) | Any picture | $0.001 |
face_recognition (Face recognition) | A picture of a person | $0.001 for your own collection and $0.001 for celebrities |
ocr (OCR) | Any picture | $0.0015 |
bynn_depth (Depth Vision) | Any picture | free |
bynn_segment (Segment Vision) | Any picture | $0.001 |
knowledge (Knowledge) | One of your knowledge bases, no picture | $0.001 a request |
- Billed: each booster and image read with status
ok, counted per booster inevidence.<name>.billed_callsand in all inusage.boosters. - Free: a booster that timed out or failed, a request the safety check refused, and a
402, which comes before any booster runs. - Billed every time: a reading is billed each time a request uses it, a repeat of the same picture and a retry after a
529included. An exact repeat of a whole request within the cache's window is billed in full too.
Collections#
A collection is billed for what it does: each search, each enrollment, and each item stored.
| What | Price |
|---|---|
| A search of your own collection, per picture | $0.001 |
A search of celebrities, per picture | $0.001 |
| An enrollment, per picture the face model reads, a face found or not | $0.001 |
| Storage, per item and month | $0.00001 |
- Searches are
face_recognitionbooster calls, billed per picture and collection searched with the request's tokens; a booster that timed out or failed is free. - An enrollment is billed once the face model has read its picture, a face found or not; a picture that cannot be read, and an enrollment refused before the picture is read, are not billed. See Collections.
- Storage is a monthly fee: every item stored at any time in a month pays that month, an item deleted during it included, with no proration. It is charged on the 1st of the next month, from the balance or on a dedicated workspace's invoice. The public collection carries no storage fee.
- Face collections are in Pro and Business, and in the trial, where searches of them and of
celebritiesare billed booster calls; on Starter acelebritiessearch runs only as a test call. An enrollment and storage are always billed. - An unpaid storage charge is retried 3 times, 10 days apart, with an email to the workspace's owner and admins each time. After the third failure every item of the workspace's collections is marked for deletion 30 days later; a top-up or a new payment method in that time retries at once, and a success clears the marks. At the date the items are deleted for good.
Knowledge#
Knowledge is billed for what it does: each request that reads a knowledge base, the new text you index, and the passages you keep.
| What | Price |
|---|---|
| A knowledge call: one request that names the booster, or one search | $0.001 |
| Indexing, per million new tokens of your documents | $0.10 |
| Storage, per passage and month | $0.00002 |
- A call is billed once per request that reads a knowledge base with status
ok, pinned documents only included; one that timed out, failed or found nothing ready is free. The passages' tokens are input tokens at the model's price. - Indexing is billed for the tokens of new text only; the same text again, and changes to a pin, a reference or metadata, are free.
- Storage is a monthly fee per passage (chunk), billed on the same monthly storage bill as collections: every document stored at any time in a month pays for the passages it last held that month.
- Kept until you delete them: a knowledge base and its documents stay stored, and billed, until you delete them, also in a workspace with zero retention: storing a document is your own instruction, separate from how long requests are kept.
- An unpaid storage bill is retried 3 times, 10 days apart; after the third failure every stored item of the workspace's collections, documents included, is marked for deletion 30 days later and then deleted, as for face collections.
- Knowledge bases are in Pro and Business, and in the trial; on Starter a knowledge call runs only as a test call, on the public sample knowledge bases. See Knowledge pricing.
Fine-tuning#
| Item | Price |
|---|---|
A training run that ends ready | 3 × its compute time at the compute price + $50 |
A run whose quality gate says no_gain | $50 |
A run that ends failed or refused | Nothing |
| A run you cancel, or whose model you delete while it runs | 3 × the compute used so far, without the $50; nothing if it had not started |
| Requests to a fine-tuned DecisionNode | $0.062 per million input tokens ($0.042 + $0.02) |
| Requests to a fine-tuned DecisionNode Flash | $0.055 per million input tokens ($0.035 + $0.02) |
| Storage and deploying | Free |
The console shows a run's estimate before it starts, and the fee is charged from the balance when the run ends, from the time it took. A run does not start while the balance is below its estimate. See Fine-tuning limits and pricing.
Number questions#
A number question is billed like a choice with one option per grid value. A count on 0 to 50 is billed like a choice of 51 options, and the default grid of 0 to 255 like one of 256. A tight grid costs less. A grid of more than 256 values is composed in the same call: it is cut into the fewest equal pages of at most 256 values (300 values make two pages of 150), one choice picks the page and one choice per page picks the value within it. It is billed for every question it is composed of, the page choice and every page, so a wider or finer grid costs more. usage.input_tokens in the response is always the exact figure. Images are priced as on every call.
Points and boxes#
Point and box questions: 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. usage.input_tokens in the response is always the exact figure.
Sessions#
A session bills its context once, when it opens: the instructions, the state and the questions, at the price per input token in force when it opens. After that, each answered frame is billed for its own tokens and the questions it asks, at that same price, whatever changes during the session. The context is never billed again per frame, dropped frames and frames that got an error are not billed (except a frame the safety check refused: the model read it, so it is billed like an answered one), and output stays free. The open response and every reply report what they billed in usage.
session = opening tokens × price
+ sum(frame tokens × price)
- opening tokens
- the instructions, the state and the questions, once
- frame tokens
- each frame's own tokens plus the questions it asks
Worked examples#
Each workload asks three questions per request. Tokens is usage.input_tokens for one request, as the response reports it.
| Workload | Tokens | Requests a month | DecisionNode | DecisionNode Flash |
|---|---|---|---|---|
| Support tickets | 62 | 1,000,000 | $2.60 | $2.17 |
| Comment moderation | 94 | 30,000,000 | $118.44 | $98.70 |
A receipt photo check of 210 text tokens and one 1,024 x 1,024 photo (about 1,026 image tokens), 200,000 times a month, costs $1.76 for the text and $18.47 for the photos on DecisionNode, and $1.47 and $16.42 on DecisionNode Flash.
Prepaid balance#
Credit tops up a plan: it is drawn once the month's allowance is used. A workspace without a plan cannot buy credit (409 plan_required); choose a plan first.
- Top up in the console under Billing, from $5 to $5,000 per payment; add more in a second payment. The balance is shared by the whole workspace, and requests draw it down token by token.
- Auto reload has two settings: a threshold and an amount. When the balance drops below the threshold, the amount is charged to your card and added; the amount must be larger than the threshold. With auto reload off you top up by hand, and requests fail once the balance reaches $0.
- One receipt per payment. Every top-up, by hand or by auto reload, gets a receipt you can download from the console. Prepaid usage is drawn from the balance, so a prepaid workspace gets no monthly bills; invoicing applies only where an Order Form provides for it.
- At zero, requests return
402Out of credit until you add credit, so with auto reload off a runaway job stops at a zero balance. A refused request is not billed, and neither is a safety check refusal. - Machine payments: an agent buys credit without a person at
POST /v1/credit, $5 to $5,000 a purchase, with a card through a Stripe Shared Payment Token, USDC on Tempo or USDC on Base; the402itself carries the challenge. The credit is booked as a top-up, with a receipt, and expires 12 months after purchase like any other.
Credit, refunds and disputes#
- Credit is prepaid for use of the API: it is not transferable and has no cash value.
- Closing your account: unused credit is not refunded when you close it, except where the law requires a refund. Deleting a workspace forfeits its balance in the same way, so spend down or move work first.
- If we end the agreement on 30 days' notice, we refund your unused credit.
- Disputing a charge: write to legal@bynn.com within 60 days of the charge.
- Credit expires 12 months after purchase; free or promotional credit expires as stated when it is granted. Credit is used oldest first.
- The binding rules are in the Terms of Service.
Watching spend#
The console's Usage page charts spend, tokens and requests by model and by key, hourly or daily (Console, keys and members has the filters). In code, read usage.input_tokens from each response; it is exactly what was billed. What happens to the data you send is on Data and privacy.