WaveSpeedAI

WaveSpeedAI does not sell plans. The pricing page says you add credits and pay per output, with no monthly fee or commitment, and each model has its own price. At the lowest resolution or quality, starting prices run from $0.005 an image (Z-Image Turbo) to $0.14 (Nano Banana Pro), and from $0.05 a second of video (Wan 3.0) to $0.18 (Seedance 2.5). Language models are priced per 1M tokens, for example Claude Opus 5.5 at $4 input and $20 output. The page says discounts may apply: the homepage shows some featured models with a struck-through price and a lower one, such as 10% off a Seedance 2.5 card, without saying which resolution or length the card price is for.

Because there is no subscription, the arithmetic is simple. At starting list prices, $20 buys about 444 Seedream 5.0 Pro images, about 166 seconds of Seedance 2.0, or about 400 seconds of Wan 3.0. Those figures assume the lowest settings; the pricing doc says resolution, duration and batch size raise the cost, and that the server recalculates the charge on submission, so the estimate on the Run button is the number to trust.

The other thing a top-up changes is your limit. New accounts start at Bronze, which the account-levels doc limits to 5 predictions a minute and 2 concurrent tasks. Any successful top-up below $1,000 makes the account Silver (500 a minute, 300 concurrent), one of $1,000 to $4,999 makes it Gold, and one of $5,000 or more makes it Ultra. The level follows a single top-up, not total spend, and the pages do not publish a minimum top-up. For a product that sends parallel requests, Bronze is a testing tier, not a production one.

Three terms decide whether the platform fits paid work. First, commercial use: the Commercial Use Policy says most models allow it, each model has its own terms, and research-only models should not be used commercially. Second, retention: generated files generally expire within 7 days, so download or collect them by webhook. Third, data: the Terms of Service, updated Sep 28, 2026, say you own your inputs and outputs and license them for providing the service, and let WaveSpeedAI use aggregated, anonymized usage data to improve it. Neither the Terms nor the Privacy Policy says whether inputs or outputs train models, so ask support for that in writing if it matters.

Refunds are narrower than they sound. Failed requests, system errors and system timeouts are credited back automatically, and unused credits never expire, but the Terms call purchased credits non-refundable. If your own client or proxy times out in Sync Mode while the task finishes, the charge stands; fetch the result by prediction ID.

1. Sign up and note whether the $1 trial credit appears and which models it can run.
2. Pick one model you plan to use, read the estimate on the Run button, and run one short job at the lowest setting.
3. Compare the credits charged with the estimate and with the pricing page’s starting price.
4. Open the terms shown for that model and note whether commercial use is stated.
5. Download the output, then check the History page after 7 days to see what remains.

Accept it if the charge matches the estimate, the trial covers a real test and the model’s terms allow your use. Reject it if the trial cannot run your model or the charge differs from the estimate without explanation. No quality, speed or cost result is recorded here.