Answer first: Wan 2.2’s open weights remove a model-license fee, not the cost of generation. Self-host when infrastructure control and sustained utilization justify operations; use an API for metered integration; use a hosted UI for evaluation or irregular creative work.
Source check: The official Wan 2.2 repository, install guide, license, and Wan-AI model collection were rechecked on August 13, 2026. Reconfirm the exact checkpoint, commit, hardware path, provider price, and hosted-plan terms before estimating cost.
Three Access Models
| Access | You pay for | Best first candidate when |
|---|---|---|
| Self-hosted open weights | GPU, electricity or cloud, storage, setup, security, monitoring, and review | Control, privacy, customization, or sustained volume matters |
| Managed API | Successful calls, duration or output settings, retries, storage, and integration | Demand is variable and programmatic access matters |
| Hosted UI | Subscription or credits, queue, plan limits, and manual operation | You are evaluating fit or generating irregularly |
Do not compare a self-hosted model download with the list price of one API call. Compare complete accepted-shot cost.
Start With the Official Wan 2.2 Repository
The current official repository documents separate text-to-video, image-to-video, text-image-to-video, speech-to-video, and animation checkpoints. Hardware needs differ materially by task and checkpoint. The repository’s example for a large A14B single-GPU task calls for substantially more memory than a typical consumer card, while its smaller 5B workflow is positioned differently.
Confirm:
- exact repository, commit, and checkpoint;
- task and resolution;
- GPU memory and runtime from your own test;
- dependency and security maintenance;
- prompt-extension dependencies;
- storage and egress;
- license and NOTICE obligations.
Self-Hosting Cost
Monthly self-host cost includes:
- amortized hardware or cloud GPU time;
- idle capacity;
- power and cooling;
- model and output storage;
- engineering setup and upgrades;
- monitoring and failed jobs;
- human review and post-production.
Divide that total by accepted shots, not submitted jobs.
Break-Even Worksheet
Use one accepted-shot unit across all three access models:
| Cost input | Self-host | Managed API | Hosted UI |
|---|---|---|---|
| Direct generation | GPU runtime, electricity, or cloud time | Model-specific billed usage | Subscription or credits consumed |
| Failed attempts | Compute plus queue occupancy | Billed retries and failed-job policy | Credits, queue time, and manual retries |
| Operations | Deployment, upgrades, monitoring, security | Integration, webhooks, fallbacks, storage | Account administration and manual export |
| Idle capacity | Hardware or reserved GPU not producing | Usually none beyond prepaid balance | Unused subscription allowance |
| Review and finishing | Human review, editing, and delivery | Human review, editing, and delivery | Human review, editing, and delivery |
Then calculate:
Accepted-shot cost = all direct, retry, operations, and review cost ÷ approved shots
For self-hosting, calculate a break-even volume only after a real benchmark. Divide the monthly fixed self-host cost by the difference between observed API and self-host variable cost per accepted shot. If that difference is zero or negative, volume alone does not justify self-hosting.
API and Hosted Cost
For an API, run a bounded sample and capture the current model, duration, resolution, credits or price, failed-call policy, concurrency, and data retention. For a hosted UI, also include subscription commitment, queue time, manual labor, watermark, and export limits.
Recheck prices at purchase time. Third-party providers can package the same Wan model with different defaults, versions, and post-processing.
License and Output Rights
Wan 2.2’s official repository states that its models use Apache 2.0 and that the project claims no rights over generated content. It also makes users responsible for lawful and non-harmful use.
That does not create rights to unlicensed source images, people, voices, trademarks, or copyrighted characters. Preserve model, input, consent, and output-review evidence for commercial work.
Related Flowith Guides
- Use Wan AI architecture and hardware FAQ before self-hosting.
- Use Wan AI alternatives when another deployment model may fit.
- Use Wan AI vs. Kling for a same-brief motion test.
FAQ
Is Wan AI free?
Wan 2.2 code and models are published under Apache 2.0, but self-hosting still has hardware, cloud, electricity, storage, engineering, and review costs.
Which Wan 2.2 model should I self-host?
Choose from the exact tasks and checkpoints documented in the official repository, then confirm GPU memory, runtime, output requirements, and license before deployment.
Is a Wan API cheaper than self-hosting?
It can be for irregular or low-volume use. Compare current API charges and retries with amortized compute, operations, engineering, and idle capacity.
Can Wan 2.2 output be used commercially?
The official repository uses Apache 2.0 and says it claims no rights over generated content, but users remain responsible for inputs, applicable law, downstream rights, and harmful or misleading use.
How should I compare Wan API and hosted-provider prices?
Run the same accepted-shot test and record the exact model version, duration, resolution, retries, failed-job policy, concurrency, storage, retention, watermark, and post-processing. A provider’s headline credit or call price is not a comparable unit by itself.
Official Sources
Sources last checked August 13, 2026: