LLMBOTTLENECK.COM
API pricing

Free to build on. Pay when it is in front of customers.

Every endpoint answers on every plan, with the same data and the same evidence. The free plan is paid for with a visible credit; Pro and Business remove it, put the widget on your site without our branding, and fit production volume.

FREE

$0 free forever

Prototypes, personal projects, open source, and anything happy to credit the data.

200calls a day
10per batch
  • Every endpoint, full fidelity
  • Credit shown next to results you display
  • The widget, with our branding
BUSINESS

$79 a month

Shops, PC builders and publishers putting the widget or the data across a catalogue.

100,000calls a day
250per batch
  • Everything in Pro
  • The widget unbranded on 10 domains
  • Your own button in the widget — a store or affiliate link, filled with the model and device on screen
  • Priority email support; a reply within one business day
  • Secure checkout by Stripe — card details never reach this site
  • Card, Apple Pay, Google Pay or Link · VAT invoices
  • Cancel any time; the plan runs to the end of the period

ENTERPRISE

Custom by contract

GPU clouds, marketplaces and platforms that need terms rather than a list price.

No capby contract
500per batch
  • Your own devices added to the catalogue
  • Response time agreed in writing
  • Invoice and purchase-order billing
Ask for a contract →

Nothing to wait for: start on Business today; the contract replaces the subscription when it is signed, and the keys stay the same.

How paying works

  1. Choose a plan and subscribe

    One button. There is no account to create first and no form of ours to fill in.

  2. Pay on Stripe’s checkout

    Card, Apple Pay, Google Pay or Link. Tax is calculated from your address and shown before you pay; add a VAT number for a business invoice.

  3. Get your key straight away

    Stripe returns you here, the payment is confirmed with Stripe, and your key is shown once — the plan is already active on it.

  • Card details go to Stripe and never reach this site; only a digest of your key is stored.
  • Cancel any time in the billing portal; the plan runs to the end of the period you paid for.
  • No overage charges: past the allowance, calls are refused, never billed.
  • Upgrades are prorated; a price you subscribed at is kept while the subscription lasts.
  • Already have a free key? Upgrade it and your code keeps working unchanged.

Which plan fits?

Pro — About 600 distinct questions a day fits Pro.

Who uses it for what

Shops and PC builders

The widget on each graphics card or laptop page answers “will it run the model I want?” in the buyer’s terms, with your own buy button. Build the tag →

Publishers and review sites

Tables and buying guides that stay correct as models ship: one batch call for the table, and /v1/changes to know when to refresh it.

GPU clouds and marketplaces

/v1/recommend and /v1/target turn “I want to run this model this fast” into the instance to rent, with the reasoning attached.

Apps and developer tools

A fit check before a multi-gigabyte download, with the offload split and a speed estimate. One call per question, typed clients for TypeScript and Python. Read the API →

Shipping the data inside your app? The data licence

An app that tells its users what fits before a download does not call an API per user — it ships the table and refreshes it with each release. The open dataset allows that with a credit; the licence removes the credit and adds the key and the support a product needs.

$990 a year, one product

  • Every published dataset version, redistributed inside one named product, with no attribution requirement
  • A Business API key for the year: 100,000 calls a day, the unbranded widget on 10 domains
  • An email the day a new version is published, with what changed
  • Priority email support, a reply within one business day

The facts restated from publishers and manufacturers are theirs, under whatever terms they publish; the licence covers this site's compilation, computed columns and evidence labelling.

Ask for the licence →See the dataset first

What every plan includes

The API
  • All endpoints: analyze, batch, recommend, target, cluster, fine-tuning, catalogue and change feed
  • Every figure with its evidence tier and its source
  • Refusals that name the missing field, never a guessed number
  • Typed clients for TypeScript and Python, generated from the live specification
  • Usage dashboard, and a warning header before the daily ceiling
The data
  • 327 models, each from a configuration pinned to its commit
  • 135 devices, each figure quoted from the manufacturer
  • Updated as models ship; /v1/changes says what is new since your last look
The widget and the site
  • One script tag, sandboxed iframe, no key to leak
  • The calculator, free, with nothing to sign up for
  • CSV export and a print sheet for PDF

Questions before buying

Why does the free plan need a credit?
Because that is how the free plan is paid for. A visible line — Hardware compatibility data: LLM Bottleneck, linked to this site — next to results you show publicly is how people who need the data find it. Internal tools and prototypes that nobody else sees need nothing. A paid plan removes the requirement.
Does paying change the answers?
No. Every plan gets every endpoint, the same data, the same evidence and the same refusals. A plan changes what you may do with the answers and how many you get a day — never what the engine says.
Should I cache responses?
Yes, and please do. An answer only changes when the catalogue does. Keep what you asked, and poll GET /v1/changes weekly to learn which models or devices moved. Most integrations need far fewer calls than they first expect.
What happens if I go over?
The call is refused with a 429 that says which ceiling you hit, when it resets (00:00 UTC) and what the next plan allows. Nothing is charged automatically and there are no overage fees. Past 80% of the day every response carries a warning header first.
Can I switch plans, pay yearly, or cancel?
Yes. An upgrade is charged the prorated difference at once; a downgrade applies now and credits the unused time on the next invoice. Yearly is ten months' price for twelve. Cancel from Usage & keys, or from Stripe's billing login without any key.
What happens after I pay?
Stripe returns you to a page on this site that confirms the payment with Stripe and shows your API key once, with a copy button and a .env download. The plan is active on that page — nothing waits for an email. If you close it before copying the key, reopen the same page within a day to replace it; the lost one stops working.
Refunds, receipts and who is selling
Stripe emails a receipt for every payment, and every invoice is in the billing portal as a PDF with your company name and VAT number if you gave them. Duplicate or mistaken charges are refunded, and statutory withdrawal rights are honoured; a voluntary cancellation runs to the end of the paid period rather than being refunded pro rata. The seller is named in the terms.
What does the widget's call to action do?
On Business, the widget on your registered domains shows your button in place of our links — a store search, a product page or an affiliate link. You may use {model}, {hardware} and {quantization} in the address, and the widget fills them in with whatever the reader has selected.

The small print, in full

  • Calls are counted per account per UTC day and reset at 00:00 UTC each day; unused allowance does not carry over.
  • Costs a call: every analysis the engine ran, whether its result is ok or unsupported (an unsupported answer is a served answer: it names the missing evidence); every configuration a batch item analysed.
  • Costs nothing: authentication failures (401); burst refusals (429 limitedBy=burst); malformed bodies and invalid input (400, 413, 422 INVALID_INPUT), including invalid batch items; daily refusals (429 limitedBy=daily); server failures while computing (the reservation is refunded, or recorded as owed when the refund itself cannot be applied).
  • A batch reserves all of its valid items at once before any runs: either the whole batch fits in what is left today, or nothing runs and nothing is counted.
  • Plans belong to the account, not to one key. A plan is in force while the subscription is active, trialing, past due (while the card is retried), legacy_unreconciled; it ends when it is unpaid, paused, cancelled, or its first payment never completes.
  • Support response times are commitments made by one person, stated as such; Enterprise terms are agreed in writing, not assumed from this page.
  • Prices are in US dollars. Tax is shown at checkout where it applies.

Answers are estimates on every plan; a paid plan buys use and volume, not a different or more certain answer. Terms for paid plans. The changelog records every change to what is sold.

llmbottleneck
catalogue 2026-10-03models 327devices 135