Pricing
Cyclotron is free. We'd be thrilled to help you use it.
Cyclotron is open source, and you can run it yourself. Managed support keeps cyclotrons aligned, meets with you every two weeks after onboarding, and makes approved changes in a dedicated AWS account. Integrated and managed work adds ongoing work at the touch-point in your product. Every plan is bring your own key: model usage is billed to you by your providers, never through us. The service fee pays for the people and agreed scope, not tokens. At list price, Jev makes about 445 decisions per cent in the cited measurement.
Free
- The full source code, under the MIT license
- You host, operate, and align it
- No account needed
$500/month
First cyclotron. Each additional cyclotron $200/month. One-time onboarding $1,500.
- Onboarding: we set up a dedicated AWS account you own, put your first cyclotron live, and train your reviewers
- Weekly check-ins for the first eight weeks, then every two weeks
- A shared Slack channel with support on the agreed service cadence
- White glove: you ask, we make the change, you approve
- A monthly alignment review and report for each cyclotron
- Upgrades, monitoring, and backups in your account
- Bring your own key: model usage is billed to you by the provider, never through us
From $2,500/month
Pilot quoted after you discuss your workflow
- Everything in Managed, bring your own key included
- We integrate and maintain the code at the touch-point in your product
- A named customer success lead and a quarterly strategy review
- Priority response in Slack
- A fixed-price pilot that ends with one cyclotron live
Quoted
- More than 5 cyclotrons in production
- We scope and quote each project
What the model calls cost
This is recorded model cost at list prices, paid to your model provider. It is not an Anthus service price, and none of the plans above include it.
Accuracy within a few points. Price apart by 46 times.
| Model | Cost per 1,000 | Cost for 400 stories | Accuracy |
|---|---|---|---|
| Jev 1.13.0 | $0.045 | $0.051 | 79.5% |
| gpt-4.1-mini* | $0.39 | $0.45 | 83%* |
| gpt-4.1 | $2.04 | $3.74 | 78% |
* gpt-4.1-mini: this recording scored 83 percent; its five recorded runs averaged 78 percent (75 to 83).
One corpus, one run per arm, same optimizer; list prices applied to recorded usage.
These figures come from an earlier recording of the same task on 400 stories, with the earlier label set. The label sets differ, so they cannot be read against the home page example.
The monthly alignment review
Every month, for each cyclotron in production:
- Audit scores
- We check how well the cyclotron agrees with your team's recent labels. The report states the uncertainty.
- Improvement round
- We review what the LLM optimizer proposed and how each candidate measured on development labels.
- Promotions
- We decide which versions go live and record why.
- Human review rate
- We adjust how many items still go to your reviewers. A random sample of items still goes to them for audit.
- Report
- You get a short written report. You can check it against the recorded run.
Premium support. Faster responses, a named contact on our team, and extended hours are available on request. Ask about premium support
Questions buyers ask first
- Where does it run, and where do stored records live?
- The managed deployment stores Cyclotron items, labels, explanations, versions, and run recordings in a dedicated AWS account that you own. We set it up with you, inside your organization. If you have no AWS at all, we can run a dedicated account for you at the same monthly price and pass AWS costs through at cost. A hosted provider receives the decision-model or LLM requests you send through your own provider account, under that provider's data handling terms.
- It runs in my account. Why is there a monthly fee?
- Because you are not paying for hosting. You are paying for the people who keep it aligned with your team, meet with you every two weeks, make the changes you ask for, and answer in Slack on the stated service cadence. If you have that engineer, fork it and run it yourself.
- Who pays for model calls?
- You do, directly. Cyclotron is bring your own key: you hold the accounts with your decision-model and LLM providers, and they bill you for usage. And it is small: at Jev's list price, 1,800 decisions in our Hard-Decisions benchmark cost four cents. The 280-story editorial run on this page cost about $0.06 in model calls. Nothing is resold through us, and our fee never includes tokens. Which provider, and whether a hosted model, an open model you run, or a small model you own, is a cost decision we help you make with the numbers in front of you. AWS bills you for the resources in your account.
- What if we need more?
- Each additional cyclotron is $200 a month, set up by us. Beyond five cyclotrons, or for improvement work beyond the review cadence, we quote custom work.
- Who owns the integration?
- The current service information describes integration and maintenance, not intellectual-property, access, or exit terms. We discuss those terms before work begins; this page does not set them.
- Which models can I use?
- Any decision model or LLM with an adapter: Jev, Kev, Laya and System One clients for decisions; OpenAI and LiteLLM for the optimizer. Another provider needs one small adapter class.
- What is one cyclotron?
- One decision about one kind of item, for example publish or reject for a story. A cyclotron can have several classifiers, each a decision that people review, and each classifier can use supporting questions. The pricing counts cyclotrons, not classifiers or questions, because what costs us attention is setting up and aligning each decision.