A generated hero scoreboard titled 'Kling Image 4.0 vs FLUX 3 Image — the scoreboard'. Left column 'Kling Image 4.0': Elo 1179.93, Rank #7, Max native 8K, References billable, Price 16 units at 2K, Strength polish. Right column 'FLUX 3 Image': Elo 1114.55, Rank #21, Max native 4K, References 10 included, Price 1.5 cents, Strength placement. Footer reads 'Elo figures per Artificial Analysis; pricing per each vendor's own rate card.'
Guides & Insights

Kling Image 4.0 vs FLUX 3 Image: Placement Versus Polish

Author

Gideon Frost

Date Published

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

Ask Kling Image 4.0 and FLUX 3 Image for the same poster and you get two different kinds of answer. Kling Image 4.0, shipped on 17 November 2026, wants one rich prompt and returns a finished-looking frame — or a whole series of them with a shared look. FLUX 3 Image, part of Black Forest Labs' FLUX 3 family that began rolling out in July 2026 and was completed by FLUX 3 Image Klein on 15 October, wants you to draw boxes: an element table with an id and a bounding-box coordinate for every object that matters, then a one-line scene prompt that ties them together. One model sells you the composition; the other lets you specify it. On Artificial Analysis's leaderboards Kling is ahead on quality, by 65 Elo points in text-to-image and 64 in editing, and FLUX is the cheaper call by 1.5 cents a frame. Neither gap explains which one belongs in your pipeline, because the models disagree about what a request is.

The two numbers that decide this

Both models appear on Artificial Analysis's text-to-image and image-editing boards, which makes this matchup one of the few where you can compare on one instrument rather than stitching together two vendors' slide decks.

• Text-to-image Elo — Kling Image 4.0 1,179.93 at 7th of 116; FLUX 3 Image 1,114.55 at 21st

• Image-editing Elo — Kling Image 4.0 1,182.89 at 6th of 86; FLUX 3 Image 1,118.97 at 22nd

• Price per image on the index — Kling 6.3 cents; FLUX 3 Image 4.8 cents; FLUX 3 Image Klein 0.9 cents

• Maximum output — Kling 8K; FLUX 3 Image 4K, which BFL demonstrates at 5456 × 3072, about 16.8 megapixels

• References per request — Kling bills each one at 0.8 units; FLUX 3 Image accepts up to 10

• Layout control — Kling: none published; FLUX 3 Image: bounding boxes with editable element tables

A screenshot of the Artificial Analysis AA-Image-T2I v2.1 leaderboard in English, showing a ranked list in which Kling Image 4.0 is seventh at 1,180 with 5,550 samples released November 2026 at $63 per thousand images, FLUX 3 Image is twenty-first at 1,115 with 6,751 samples released October 2026 at $48 per thousand images, and the GPT Image 2.6 and 2.5 tiers occupy the top positions.

The 65-point gap is real but it is one instrument on one day, and this blog has watched index revisions move scores by double digits. Treat it as "Kling is preferred more often in blind comparisons" rather than "Kling is 6 percent better," because the Elo scale is not a percentage of anything.

Bounding boxes are the whole FLUX argument

Black Forest Labs built FLUX 3 Image around a 0-to-1000 coordinate grid on both axes. You define elements — each with a name, a box, and a description — and the model renders inside the boxes. The demo on BFL's own model page is a five-element poster: a text element at [10, 200, 170, 800] reading "LE FESTIVAL DU SOLEIL" in a thin serif, a town, a dome, swimmers, a crowd, plus one line of scene prompt. Elements you do not touch stay locked; that is what makes the model usable for revisions rather than one-shot generation.

For agent work this is the more interesting design. BFL documents a loop where an LLM writes the caption and the element table, the boxes stay editable, and the model generates inside the layout the agent planned. An agent can therefore produce a layout a human can inspect — a table of boxes and descriptions — rather than a prompt nobody can debug.

A generated diagram titled 'Two models, two ideas of a request'. Left card 'Kling Image 4.0' shows a single prompt box leading to a finished frame or a series with a shared look, noted as having no published layout control. Right card 'FLUX 3 Image' shows a three-row element table with bounding-box icons for a woman, a logo and a headline leading to a one-line scene prompt, noted as using bounding boxes and an editable element table. Footer reads 'Layout interface per each vendor's own documentation; board positions per Artificial Analysis.'

Kling's approach is the opposite bet. Image 4.0's "coherent image sets" feature takes one prompt and returns a series with a consistent look, and its "long image generation and editing" mode produces tall, page-shaped output — a full webpage layout at 140 units, or an e-commerce detail page at 60 to 220 units depending on length. Both are composition problems solved by the model rather than by you. If your brief is "twelve consistent product frames," Kling's framing is less work. If your brief is "this logo here, this headline there, this product at 60 percent width," FLUX is the only one of the two that takes the instruction.

Resolution and the reference bill

Kling goes to 8K natively and charges for the climb: 16 units for a 2K still, 24 for 4K, 40 for 8K, against a package unit price of $0.00384 at the $300 tier. That is roughly 6.1, 9.2 and 15.4 cents. FLUX 3 Image renders natively at 2K and 4K — BFL's example is 5456 × 3072 — and the index lists it at 4.8 cents. On a like-for-like 4K render the two are close, with FLUX slightly cheaper; at 8K Kling is competing with nobody in this pair.

The reference economics cut the other way. Kling made references billable with this release: 0.8 units per reference image or reference element, and a reference element counts the same as an image. FLUX 3 Image takes up to 10 references and, as far as its published material says, does not meter them separately — you pay per output. A ten-reference composite therefore costs Kling 8 units of input on top of the output, about 3 cents, and costs FLUX nothing extra. At one image that is noise. Across a thousand-asset catalogue it is a line item that appears out of nowhere, which is exactly the kind of change this blog exists to flag.

A screenshot of Kling's Open Platform pricing and purchase page in English, showing the Image API section with 2K, 4K and 8K output prices in units, the 0.8-unit reference-image and reference-element charge flagged as newly billable, and long-image prices of 140 units at 1K and 152 at 2K, with prepaid package unit prices near four-tenths of a cent.

The cheap tier changes the question

FLUX 3 Image Klein is the model most teams will actually run at volume. It is a 5.6-billion-parameter open-weights release from 15 October 2026, it scores 1,041.34 on text-to-image and 1,068.50 on editing, and it lists at nine-tenths of a cent an image — a fifth of FLUX 3 Image and a seventh of Kling Image 4.0. BFL says Klein leads its own blind evaluations for text-to-image, multi-reference generation and editing, which is a vendor statement about a vendor's own family; the independent index puts it 31st and 36th respectively, well behind its bigger sibling. Still, for draft passes, layout exploration and anything a human will retouch, a ninth of a cent per frame changes how you budget experimentation.

A sensible split for a catalogue team: Klein for exploration, FLUX 3 Image or Kling Image 4.0 for final frames, and 8K only for the assets that will be printed or zoomed. That is not a compromise position — it is what the price ladder and the Elo spread are telling you to do.

Running both from one place

The practical snag with a two-vendor comparison is that it normally means two contracts, two billing relationships and two sets of rate limits. OrcaRouter's pitch is that it should not: one API key across 200-plus model identifiers with provider list pricing passed through at 0% markup, so a Black Forest Labs or O​penAI price change shows up on your invoice the same day it happens rather than at the next contract renewal. Automatic failover matters more than usual with image models, because a provider-side timeout mid-batch is indistinguishable from a bad prompt until you retry. Kling Image 4.0 is not in our catalogue — that one is a direct Kling API call — but the rest of this comparison table, including the gpt-image and Imagen families, is one key away, which is enough to settle the layout question before you commit to a vendor.

Which one, concretely

Pick FLUX 3 Image when the composition is the deliverable: templated marketing layouts, anything with a locked logo position, agent-generated designs you need to audit, and multi-reference composites where ten source images have to land in specific places. The bounding-box interface is the only mechanism in this pair that lets a human or an agent state where things go, and for template work that is worth more than 65 Elo points.

Pick Kling Image 4.0 when the look is the deliverable: a coherent set of twelve campaign frames, a tall page-shaped graphic, or a single hero image at 8K where the small text has to survive. Its editing score is the higher of the two by a wide margin, and the set and long-image modes have no direct FLUX equivalent.

Pick FLUX 3 Image Klein for everything you intend to throw away. At nine-tenths of a cent a frame, the correct number of drafts is "more."

And test before you commit: give both the same brief with a logo that must sit bottom-left at a fixed size, and see which one obeys. That single test will tell you more about fit than either leaderboard row.