A generated hero card titled 'Kling Image 4.0: 8K, and a New Reference Bill', dated 17 November 2026, with rows for the 2K, 4K and 8K output tiers and a highlighted row reading '0.8 units' under the label reference images, and a footer reading 'Kling pricing per Kling's developer rate card; Elo figures per Artificial Analysis.'
Guides & Insights

Kling Image 4.0 Ships 8K Stills — and Stops Giving References Away

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Kling Image 4.0 went live on 17 November 2026, and the headline Kuaishou chose is about pixels: native 8K stills, better small text, and coherent image sets from a single prompt. The detail that will change budgets is smaller and buried in the developer price sheet. For the first time in the Kling image line, reference images and reference elements are billed inputs. Kling's own documentation says so in one line — both are billable now, "which differs from previous models, where reference images and reference elements were free." If your workflow is twelve product shots per SKU with three references each, that sentence costs more than the jump from Kling Image 3.0 Omni to Kling Image 4.0 ever will.

The launch itself is real and current: the release history on kling.ai carries a 2026-11-17 entry headlined "Kling Image 4.0 Is Here!", Kling's account posted the availability note in the early hours of 18 November, and the model already has rows on both Artificial Analysis image boards. This is not a leak, a preview, or a region-limited beta. It is a shipped model with a published unit price, which is exactly why the billing change is worth reading before the feature list.

A screenshot of Kling's release-notes page in English, headed Release Notes, with the most recent entry dated 2026-11-17 marked Major and titled 'Kling Image 4.0 Is Here!', described as stunningly realistic and sharp with results from a single image to a cohesive series, above older entries from 15 October 2026 and 17 June 2026.

What the release actually claims

Kling's product page lists five capabilities for Image 4.0, each with a one-line gloss. They are unusually specific for marketing copy, so they are worth quoting rather than paraphrasing:

• True-to-life realism — "natural lighting, lifelike results"

• Native 8K resolution — "rich detail, intricate textures"

• Exceptional text rendering — "crisp small text, creative typography"

• Coherent image sets — "one prompt, a complete series"

• Long image generation and editing — "full-page view, layered editing"

Every one of those is a vendor claim, unreproduced by us. What can be checked independently is the quality signal, and it is strong without being top of the table. On Artificial Analysis's text-to-image board Kling Image 4.0 sits seventh at an Elo of 1,179.93; on the image-editing board it is sixth at 1,182.89. Its predecessor, Kling Image 3.0 Omni, is 87th on the text-to-image board at 925.07. A 255-point jump in one generation is the single most useful number in this release, and it is not a number Kling put in the announcement.

The 8K claim, and what an Elo score cannot see

Native 8K is the claim with the least independent support, because no public benchmark scores resolution. Artificial Analysis runs fixed-size comparisons and reports a preference score; it does not measure whether a 7,680-pixel-wide render holds detail at 100 percent zoom, and Kling has not published a sample set with pixel dimensions attached. Treat "native 8K" as a specification you verify on your own artwork, not a fact about quality. The distinction matters because 8K is also the tier where the price ladder bends hardest — 40 units an image against 16 at 2K — and the claim is doing commercial work, not just technical work.

The text-rendering claim is easier to test cheaply, and it is the one most buyers care about. Kling frames it as small-text legibility, which is the failure mode that ruins packaging mockups and dense UI comps: a headline renders fine and the 8-point legal line turns to soup. Whether 4.0 fixes that is a question for your own type samples, but the claim is at least aimed at the right problem.

The reference bill is the structural change

Image 4.0 prices references on the input side of the equation. Kling's formula is total cost equals the number of input reference images and elements times the input price, plus the number of output images times the output price — with one reference element counted as one reference image, at 0.8 units each. That is a small number per reference and a large number per campaign. A twelve-shot set built from four references moves from 192 units of output to 230 units of total, roughly a fifth more, and the surcharge scales with how reference-heavy the work is rather than with how much output you get.

• Output, single image — 2K 16 units, 4K 24 units, 8K 40 units

• Output, image set — the same 16 / 24 / 40 ladder applies per image in the series

• Reference input — 0.8 units per reference image or reference element, newly billable

• Long image, webpage — 140 units at 1K, 152 units at 2K

• Long image, e-commerce detail page — image-to-image only, 60 / 120 / 200 units at 1K for short, medium and long, and 72 / 140 / 220 at 2K

• Unit price — $0.00384 in the $300 package, falling to $0.00381 at $999 and $0.00373 at $4,000

A screenshot of Kling's Open Platform pricing and purchase page in English, showing the Image API section with per-image unit costs at 2K, 4K and 8K, the 0.8-unit charge for reference images and reference elements, long-image prices of 140 units at 1K and 152 at 2K, and the e-commerce detail-page ladder of 60, 120 and 200 units at 1K and 72, 140 and 220 at 2K, with prepaid package unit prices near four-tenths of a cent.

Put those together and a 2K still is about 6.1 cents, a 4K still about 9.2 cents and an 8K still about 15.4 cents at the entry package rate. A reference-heavy 2K set costs a little over 6 cents a frame plus roughly a third of a cent per reference per frame. Artificial Analysis's listing for the model, 6.3 cents an image, sits comfortably inside that arithmetic, which is a good sign that the published unit price is the real one and not a teaser.

Coherent sets and long images are the products, not the features

The two capabilities that separate Image 4.0 from a generic text-to-image model are the ones Kling prices most elaborately. "Coherent image sets" means one prompt returns a series with a shared look — the thing agencies otherwise hold together with seeds, style references and a retoucher. "Long image generation and editing" means tall, page-shaped output: a full webpage comp at 140 units, or the e-commerce detail-page format that only accepts an image-to-image request and climbs from 60 units for a short panel to 200 for a long one.

That last row is the most interesting line in the price sheet, because it describes a workflow rather than a picture. A detail-page image is a Chinese e-commerce convention — a single tall graphic with the product, the claims, the specs and the lifestyle shot stacked in order — and Kling is selling the model as a way to produce one directly instead of compositing it. If you have ever built that layout by hand in a design tool, the medium tier at 120 units, about 46 cents, is not an expensive experiment.

Where it sits against Kling's own line

Kling Image 3.0 Omni shipped on 31 January 2026 with 2K and 4K output, image-series mode and multi-reference support, and it has been the line's reference-free workhorse since. Image 4.0 does not replace it in the catalogue — the 3.0 identifiers are still listed — and the price gap between them is not small: 3.0 Omni lists at 28 units an image against 16 for a 2K Image 4.0 still, so the newer model is cheaper per frame at the same resolution and dramatically better on the leaderboard. The migration case is unusually clean for a version bump, provided your prompts do not lean on free references, in which case the arithmetic needs redoing per asset.

One practical note on access: OrcaRouter does not currently serve Kling Image 4.0, and this piece is not going to pretend otherwise. What the catalogue does carry is a route to the image models we do host — the gpt-image-2.5 pair, Imagen, Gemini image models — on one key with provider list pricing passed through at 0% markup, which is the honest way to run a bake-off against Kling's numbers without opening a second vendor account. If Kling's 8K output is what you need, go to Kling's own API; if the question is whether your pipeline should pay 15 cents a frame, the comparison is worth running on models you can already call.

A screenshot of the OrcaRouter pricing page in English, headed 'We never mark up the tokens you buy', with a Zero Markup Guarantee panel stating that routing is free and that the rate paid is the provider's rate, and a rate table listing models with input, cached input, cache write and output prices per million tokens alongside an OrcaRouter fee column reading $0.00.

What to test first, and what to ignore

Three tests will tell you more than any launch thread. Render one 8K frame of your densest real asset and inspect it at 100 percent, because resolution claims fail at the edges, not in the centre. Render one small-text layout at 2K and read the smallest line, because that is where the typography claim either holds or does not. Then rebuild one existing reference-driven set and compare the invoice, because the reference bill is the only part of this release that is guaranteed to change your cost per asset. Everything else — the realism language, the "complete series" framing — is a claim you can confirm or dismiss in an afternoon.

The release is a genuine step for Kling's image line, and the 255-point leaderboard move is the evidence for it. The billing change is the part that will be quoted in planning meetings six months from now, and it is the part the announcement did not lead with.