CUA-S1 Pricing 2026 — Free Model Weights, Separate Fleet Costs
Official pricing and availability checked:
Verified pricing boundary
The checkpoint is open; no hosted CUA-S1 meter exists
CUA-S1-FORMS is a public 706,048-parameter checkpoint distributed under the MIT license. CUA has not announced a first-party hosted inference endpoint, token price, request fee, subscription, or model-specific credit rate. Running it locally avoids a model usage charge, but the surrounding automation still consumes compute and engineering time.
CUA-S1 model fee
$0 usage fee · MIT
Hosted model API
Not announced
Checkpoint
2.8 MB · 706,048 parameters
Canonical API rows
0
Optional Cua Fleet pricing
Cua Fleet supplies desktop infrastructure. These official rates are separate from the CUA-S1 checkpoint and do not create a hosted model API.
| Meter | Official rate | What it covers |
|---|---|---|
| Virtual CPU | $0.044625/vCPU/hour | Fleet desktop compute |
| Memory | $0.0223125/GB/hour | Fleet desktop memory |
Example: an always-on 2-vCPU, 4-GB allocation is $0.1785/hour, or about $130.31 for 730 hours before storage, networking, observability, document processing, and review. This is a mathematical illustration, not a quoted CUA-S1 bundle.
What CUA-S1 actually does
CUA-S1-FORMS is a narrow option scorer for structured form tasks. For each interface element, it chooses among values already extracted from a document plus check, click, and skip. It does not generate new values, interpret arbitrary screenshots, plan a complete workflow, or decide the final action order.
The public checkpoint uses a byte-level two-layer transformer encoder with an option-attention head. CUA reports 99.95% top-1 accuracy on a form-disjoint synthetic test and 100% on a 196-decision real demo set. Those are project-reported, narrow results—not proof of general desktop autonomy.
Why CUA-S1 is not in the pricing API
Our canonical pricing dataset and public API describe billable model endpoints. CUA-S1 has no first-party endpoint or usage unit, while Cua Fleet is general desktop infrastructure rather than a model SKU. Adding a zero-dollar model row would imply a callable free API that does not exist.
Labs inclusion decision
The exact checkpoint is downloadable, but it cannot enter the current Cost-per-Task text leaderboard. That suite expects generated text answers; CUA-S1 instead scores a fixed option set for each structured UI element. A useful test needs pinned forms and source documents, exact accessibility snapshots, a fixed driver and checkpoint, cold and warm hardware measurements, action-level and completed-form grading, abstention thresholds, retries, reviewer time, and total infrastructure cost.
Until that separate harness exists, Labs records an explicit task-contract and cost-measurement blocker. We do not convert CUA's synthetic accuracy, small real demo, or local latency claims into a leaderboard result.
Buyer guidance
- Use CUA-S1 only when the allowed values and actions can be bounded before inference.
- Measure cost per independently verified form, including failures and human escalation.
- Keep execution disabled during evaluation, then gate fill and submit actions separately.
- Route novel layouts, missing values, and consequential actions to a general model or human.
- Do not compare Fleet compute rates directly with token prices; they meter different layers.
Read the CUA-S1 launch and benchmark analysis for the architecture, published evidence, and deployment caveats.
Frequently asked questions
How much does CUA-S1 cost?
CUA-S1-FORMS is downloadable under the MIT license with no per-token or per-request model charge. You still pay for the machine, document extraction, desktop runtime, monitoring, and human review used around it.
Does CUA-S1 have a hosted API price?
No first-party hosted CUA-S1 endpoint or token rate was announced at launch. The public release is a local checkpoint and research integration, not a metered model API.
What does Cua Fleet cost?
Cua lists optional Fleet infrastructure at $0.044625/vCPU-hour and $0.0223125/GB-hour. Those are desktop-compute rates, not CUA-S1 inference prices.
Why is CUA-S1 absent from the pricing API?
The pricing API tracks billable hosted model endpoints. CUA-S1 has no public endpoint, model meter, or token price, so a zero-dollar row would misclassify local open weights as a hosted service.
Is CUA-S1 in AI Pricing Guru Labs?
Not in the text leaderboard. CUA-S1 scores choices for structured form elements rather than generating text answers, so it needs a separate form-action harness with pinned documents, interfaces, hardware, outcome grading, latency, and total workflow cost.
Official sources and maintenance
Status checked against CUA's official repository, the linked Hugging Face model card and checkpoint, the official dataset card, and Cua Fleet pricing. Sources are watched daily for license, endpoint, model-price, and Fleet-rate changes.