CUA-S1 Pricing: MIT Weights and Cua Spaces Plans
MIT local weights have no model usage fee; Cua Spaces sells sharing and relay access, not computers or CUA-S1 inference.
Source review unresolved — the table shows last-reviewed terms, not verified current offers. Check the official sources before buying.
| Product | Monthly billing | Yearly billing | Included / boundary |
|---|---|---|---|
| CUA-S1-FORMS weights | $0 model usage fee | MIT checkpoint | Local artifact; computer costs separate. No hosted inference entitlement. |
| Spaces Free | $0/month | $0/year $0/month equivalent | 2 live sessions · 20 remote hours/month |
| Spaces Pro | $18/month | $180/year $15/month equivalent | 5 live sessions · 200 remote hours/month |
| Spaces Teams | $30/month per seat | $300/year per seat $25/month equivalent | 5 live sessions · 200 remote hours/month per seat, pooled |
Bring computers you own or rent in your own cloud account. Your cloud provider bills for VMs; Cua does not host or rent the computers. These plans do not promise CUA-S1 model access or inference.
Price: Last source verification (overdue): 2026-10-25 00:06:53 UTC; access: Last source verification (overdue): 2026-10-25 00:06:53 UTC; limits: Last source verification (overdue): 2026-10-25 00:06:53 UTC; billing: Last source verification (overdue): 2026-10-25 00:06:53 UTC. These are source-review clocks, not build times.
Structured findings and source evidence · Official Spaces prices
What each Spaces plan includes
- Free: Spaces, Teleport and sharing with named people; no account needed to start.
- Pro: One person; includes Free. Expiring share links and 90-day audit history are coming soon, not included yet.
- Teams: Per-seat sessions and hours form a shared pool. Shared team computers, roles and device approval; includes Pro. SSO and audit export are not included yet.
Remote hours count each hour with at least one live relay session, per calendar month (UTC). A person or agent holding a desktop, terminal or process stream open is a live session. Local-network and direct access do not count.
No cutoff at the allowance: the app shows an upgrade notice. No per-hour overage rate is published on the reviewed pricing page.
Billing and license conditions
Monthly and annual plans auto-renew until cancellation. Yearly prices are billed upfront, not monthly; two months free. No trial. Cancel within 14 days of the first yearly charge for a full refund; no partial-month refund. Seat and interval changes are prorated by Stripe; cancellation keeps access until the paid period ends, then Free.
Teams is for business use, 1–100 seats; checkout requests company name and address, with an optional VAT/GST ID.
Spaces is source-available under FSL-1.1-MIT, becoming MIT two years after each release; it is not the MIT checkpoint license. Driver and Lume are MIT.
Former Fleet vCPU-hour and GB-hour quotes are not current prices on this page and are withdrawn from buyer comparisons.
Exact checkpoint and access boundary
cua-ai/cua-s1-forms has 706,048 parameters under MIT. The reviewed repository pins revision 4a7a9f42a3d42e6dfbd111c0c843e37ac50f1332, using safetensors + JSON.
Related tinyx checkpoint; not a formally mapped release of cua-s1-form-v0. Other Cua-S1 checkpoints have separate artifacts and licenses.
No first-party CUA-S1-FORMS inference endpoint or token/request price is published in the reviewed cards and pricing page. This is not a claim of a free callable API.
224-byte element context; 96-byte options. Selects extracted entities, check, click or skip; does not generate text or invent values.
Benchmark limitations and Labs status
The HF card attributes 99.95% synthetic accuracy and 100% on 196 real-demo decisions to the checkpoint author. The supporting results are not reproduced by libs/cua-s1.
Outside its label vocabulary it can return high-confidence skip, not reliable abstention. The current repository card reports 29.3% on 41 held-out out-of-catalogue decisions; this small external evaluation is not a general benchmark.
Not in the text leaderboard. A separate pinned form/action harness, independent completion checks and full workflow cost measurement are required.
Read the dated CUA-S1 launch analysis or the Labs coverage notes. The canonical API pricing dataset is unchanged: local weights and sharing plans are not hosted token SKUs.
Frequently asked questions
How much does CUA-S1 cost?
The exact cua-ai/cua-s1-forms checkpoint has MIT weights with no local model usage fee. Compute, extraction, automation and review still cost money.
Do Spaces plans include hosted CUA-S1 inference?
The reviewed pricing page sells sharing and relay access to your own computers. It does not promise CUA-S1 inference, rented computers or a token allowance.
Are the former Fleet rates current?
No. The current page lists Spaces Free, Pro and Teams plans instead of the former CPU and memory hourly quotes.
Can a confident skip prove a form is complete?
No. The current repository model card reports high-confidence skip on unfamiliar labels. Check label scope and completed outcomes independently.
Why is CUA-S1 absent from the API pricing table?
A local checkpoint is not a billable hosted model endpoint. A zero-dollar token row would imply a callable free API that the reviewed sources do not establish.
Official sources and maintenance
- https://raw.githubusercontent.com/trycua/cua/main/libs/cua-s1/MODEL_CARD.md
- https://huggingface.co/cua-ai/cua-s1-forms/raw/main/README.md
- https://huggingface.co/datasets/cua-ai/cua-s1-forms/raw/main/README.md
- https://cua.ai/pricing
Daily checks can reverify this reviewed payload only when sources, claims and rendering bindings agree. Changed or unavailable sources stay unresolved until reconciled. Source fetches and build times alone never refresh the four verification clocks.