WaveSpeedAI

Are Failed AI Video Generations Charged? Kling and Runway Rules Compared (2026)

Do you pay when an AI video generation fails or is blocked by moderation? What Runway's API documents, what is known about Kling, and how to stop paying for clips you never receive.

By WaveSpeedAI7 min read

Whether you pay for a failed video generation depends on why it failed. On Runway’s API, a job that fails for an internal reason is not charged, but a job rejected by input moderation is: the documentation says credits are not refunded for those. Kling does not publish an equivalent rule that we could find, and users of its app describe paying again and again for retries. In both cases the number that matters is not the price of a clip but the price of a clip you can use.

This guide sets out what each platform documents, what users report, and how to keep failed and blocked jobs from eating a budget.

Sources checked November 22, 2026: Runway’s API documentation, Kling’s public developer pages and Trustpilot page.

Three kinds of “failed”

People use one word for three different outcomes, and they are billed differently.

OutcomeWhat happenedWho usually pays
Technical failureThe job errored or timed out on the platform’s sideNormally the platform
Moderation blockYour prompt or image, or the generated video, was refused by a safety checkDepends on the platform; this is where the surprises are
Unusable resultThe job succeeded but the clip ignored the prompt, has artifacts or barely movesAlways you

Runway: what the documentation says

Runway’s API documents its task failure codes clearly.

Failure codeMeaningRetry?Credits
SAFETY.INPUT.*Your input was rejected by moderationNoNot refunded
SAFETY.OUTPUT.*The generated output was rejectedNoNot stated
INTERNAL.BAD_OUTPUT.*The model produced a bad resultYes, after revising the promptNot stated
ASSET.INVALIDInput media has wrong dimensions, duration or formatNo, fix the inputNot stated
THIRD_PARTY.UNAVAILABLEA dependency is unavailableLaterNot stated
INTERNALPlatform errorYes, after a delayNot stated

The sentence to remember is: “Unlike other failures, credits are not refunded for SAFETY.INPUT.* failures.” So a reference image that trips the filter costs the same as a finished clip. The same page advises against retrying safety failures.

Runway also limits how much you can run at once through usage tiers. The first tier allows one video generation at a time and 50 per day; higher concurrency and daily limits unlock automatically as your total purchases grow, up to 20 concurrent video jobs at the top self-serve tier.

Kling: what is and is not known

We looked for an official statement of how failed tasks are billed on Kling’s developer pages and did not find one on the public pages. If you use Kling’s own API, the reliable approach is to measure it: run a few tasks you expect to fail, then compare your balance and the per-task billing records before and after.

What is public is user feedback on the Kling app. Its Trustpilot page shows a TrustScore of 1.2 from about 450 reviews, with the usual caveat that an unclaimed profile attracts mostly unhappy users. In reviews from November 2026 the recurring themes are:

  • Paying for retries. Results with little or no motion, edits that return nearly the same video, and corrections that introduce new errors, each attempt charged.
  • Waiting. Long queues, including for users who bought credits to skip them.
  • Support. Replies that suggest better prompts with no credit offered, or no reply.

Most of these are the third kind of failure in the table above: the job completed, so it was billed, but the clip was not usable. No platform refunds that, which is why it has to be managed on your side.

What it does to your real cost

A simple way to see the effect is to compute the cost per accepted clip:

cost per accepted clip = price per generation × generations submitted ÷ clips you kept

If one in five submissions is blocked or unusable and all five are charged, your real cost is 25% above the list price. If you need three attempts per usable clip, it is three times the list price. When you compare platforms, compare this number.

How to stop paying for clips you never receive

  1. Validate inputs before you submit. Check image dimensions, file size, duration and format in your own code. An invalid asset should never reach a paid endpoint.
  2. Screen prompts and reference images first. Run user-supplied text and images through a cheap moderation check before the video job. Real people, logos, watermarks and overlaid text in reference images are the most common triggers.
  3. Never auto-retry a safety failure. Change the input or stop. An automatic retry on a billed moderation failure pays twice for the same refusal.
  4. Draft cheap, finish expensive. Find the prompt at the shortest duration and lowest resolution, then render the final at full quality once.
  5. Log every task with its failure code and its charge. Reconcile the two weekly. This is the only way to know your real failure rate and to have evidence when you ask for a credit.
  6. Show users the reason. If your own customers trigger the block, tell them the input was refused instead of showing a generic error, or they will press the button again.

A platform where failed tasks cost nothing

WaveSpeedAI offers Kling 4.0 alongside Veo, Seedance, Wan and other video models through one API, with a simpler rule for failures:

  • Failed tasks are refunded in full. If a task does not complete, for a technical reason or because a content check stopped it, 100% of its cost returns to your balance automatically.
  • The price is shown before you run, for the exact duration and resolution you chose.
  • No cold starts. Rate limits depend on your account level and are listed in the documentation.
  • One API for every model. The same two calls and the same result format work across the catalog, so trying a second model for a prompt that one model refuses takes a one-line change.
  • Support that answers through the chat on the site or on Discord.
curl --request POST 'https://api.wavespeed.ai/api/v3/kwaivgi/kling-v4.0/text-to-video' \
  --header "Authorization: Bearer $WAVESPEED_API_KEY" \
  --header 'Content-Type: application/json' \
  --data '{"prompt": "A ceramic cup of coffee on a wooden table, steam rising, slow push-in, morning light"}'
curl --request GET "https://api.wavespeed.ai/api/v3/predictions/$PREDICTION_ID/result" \
  --header "Authorization: Bearer $WAVESPEED_API_KEY"

The result call returns data.status. When it is failed, the charge has been returned; when it is completed, the video URL is in data.outputs.

One thing stays the same everywhere: a completed clip that you simply do not like is a completed task. The steps in the previous section, especially drafting at low cost, are how you keep that number down.

To try the models without code, use the video generator.

FAQ

Does Runway charge for failed generations?

For input moderation failures, yes. Runway’s API documentation says credits are not refunded for SAFETY.INPUT.* failures, unlike other failures.

Does Kling charge for failed generations?

We did not find an official public rule for the Kling API. App users report being charged for each retry when a result is unusable. Measure it on your own account by comparing the billing record of a failed task with your balance.

Why was my video blocked by moderation?

The most common causes are a recognizable real person, a brand logo or watermark, overlaid text in a reference image, or wording the filter reads as unsafe. Change the input; do not resubmit the same one.

How many videos can I generate at once on Runway’s API?

One at a time on the first usage tier, rising to 20 concurrent video jobs at the highest self-serve tier. Tiers upgrade automatically with total purchases.

Is there an API that refunds failed video generations?

Yes. WaveSpeedAI refunds failed tasks in full, including tasks stopped by a content check.

Share