# Kling Image 4.0 API: References Now Belong in Your Cost Estimate

Kling Image 4.0 bills reference inputs as well as generated images. Update your cost model, choose output counts explicitly and handle asynchronous tasks carefully.

By Paras Madan | 2026-11-18


An image-generation estimate that counts only outputs needs revisiting for Kling Image 4.0. The API now bills reference inputs too, so an application that attaches a large reference collection to every request can incur costs its old estimator never counted.

Kling's [November 17 API update](https://kling.ai/document-api/updates/api) confirms the change from output-only billing to counting input references and output images. Output rates vary by resolution. Standard images and image sets support 2K, 4K and native 8K; long images support 1K and 2K.

I could verify the billing structure, but not a current numerical rate card. Check the current official rates before quoting a customer or authorizing a batch.

## Count the assets you actually send

My first change would be to record the number of reference inputs alongside requested output count and resolution. Keep those fields in the internal estimate and the completed-job record. Do not let “one generation request” stand in for the number of billable items.

Review how your application constructs the reference list. A convenient design might attach every saved brand image, even when the request needs only a product photograph. Replace that habit with deliberate selection: each asset should have a clear purpose in the current output.

The [official Image 4.0 manual](https://kling.ai/document-api/api/usage-guide/kling-image-4-0) says assets remain effective references even when the prompt does not explicitly mention them. Removing an `@id` from the text is therefore not the same as removing the asset from the request.

For a first validation, I would submit the smallest useful reference set, inspect the output and compare the recorded charge with the estimate. Add another reference only when it solves a specific consistency problem. This is a proposed workflow, not a measured cost-saving claim.

## Choose the output type before choosing the count

The new standard endpoint is `/image-generation/kling-4.0`. Coordinated series use `settings.series_amount`, accepting two to ten images or `auto`. The manual distinguishes those coordinated sets from independent outputs controlled by `settings.n`.

I would use an explicit count for the first production rollout. Letting the model choose is convenient for exploration, but a fixed deliverable is easier to budget and review. Keep a separate allowance for revisions rather than assuming the first set will contain every usable asset.

A long page is a different product shape. `/long-image-generation/kling-4.0` produces one long image, not HTML or an editable design document. Its `settings.skill` selects `ecommerce` or `webpage`; the length setting applies to e-commerce, not webpage output.

That distinction matters before promising a finished website to a customer. Treat a generated web layout as visual material for the implementation workflow, not as deployed software.

## Track the task before considering a retry

The manual warns that successful task submission is not completed generation. Retain the returned task identifier, inspect status, and preserve `request_id` for debugging. It also advises against automatically resubmitting creation requests whose outcome is unknown.

I would make that distinction visible in the application: submitted, awaiting result, completed or failed. A network interruption should lead to investigating the existing task, not silently creating another generation. Have the job record retain the selected assets and settings so a reviewer can reconcile what was requested with what arrived.

The migration is less about adding an 8K option than making the request explicit. Count references, select the intended output form, budget revisions and reconcile completed jobs. Those checks remain useful even when the published rates change.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI teams explain API changes and creative workflows clearly enough for developers to adopt them.

[Book a 30-minute call](https://cal.com/paras-madan/30min).
