The 4K Tier Is Where Image Pricing Breaks

Black Forest Labs published FLUX 3 Image Klein on 15 October: 5.6B parameters, open weights, and a price list that looks like a mistake next to its bigger sibling. At 4K, Klein costs $0.036 per image. FLUX 3 Image, released two weeks earlier, costs $0.607. Same company, same canvas size, seventeen times the money.

Most of what got written about Klein led with speed, because a small distilled model is an easy story. The interesting part is the shape of the pricing curve, since that is what changes a budget.

Read the ladder from the top

Both models are sold per image and priced by the resolution you ask for. Both get more expensive as the canvas grows. They do not get expensive at the same rate:

• FLUX 3 Image: $0.041 at 768sq, $0.048 at 1K, $0.100 at 2K, $0.607 at 4K. Going from 1K to 4K multiplies the price by 12.6.
• Klein: $0.008 at 768sq, $0.009 at 1K, $0.011 at 1.5K, $0.014 at 2K, $0.036 at 4K. The same 1K-to-4K step multiplies by 4.

Near the bottom of the ladder the two models are about five times apart. At the top they are seventeen times apart. Anyone comparing them on a 1K demo, which is how most model comparisons get made, would file Klein as the budget option and move on. That conclusion is right for small images and wrong for large ones, because Klein's curve stays flat exactly where the full model spikes.

Put numbers on it. A hundred 4K images on Klein is $3.60. The same hundred on FLUX 3 Image is $60.70. A single Klein image at 4K ($0.036) is also cheaper than a single FLUX 3 Image at 1K ($0.048), which is a strange sentence to write and still true.

Why the curve bends

FLUX 3 Image's 4K price is not a rounding artifact. A full-quality render at sixteen megapixels asks the pipeline to denoise a far larger latent, and the vendor prices the compute it expects to spend. Klein is distilled to 16 steps at CFG 1 and carries 5.6B weights, so each 4K render asks for a fraction of that compute. The flat top of the ladder is what cheap steps buy.

Which cuts both ways. Fewer steps and a pinned guidance scale mean you have less machinery to steer with, and at CFG 1 there is no guidance knob worth turning, so direction has to live in the prompt. Whether that is a real loss depends on whether you were steering with CFG or writing better prompts.

The training story runs the other way. Klein Base is the fine-tuning variant at 50 steps and CFG 3.5, and LoRAs trained on Base load into the distilled model. So the split is straightforward: Distilled for iteration and production, Base for training runs. If you have never built one, the same LoRA dataset decisions apply here unchanged.

Same request, different endpoint

Klein keeps the FLUX 3 Image request shape. prompt, images, aspect_ratio, resolution and the bounding box format are unchanged, and switching endpoint is the migration. Up to ten reference images, 768sq up to 4K, and text-to-image and editing priced identically at every tier. An edit is not a premium here.

That makes routing mechanical rather than architectural. If your pipeline already writes FLUX 3 requests, it can address Klein by changing a string, and the only other decision is which resolution value goes in the payload.

One thing to watch when you route: reseller listings carry their own published schedule, so read the one attached to the endpoint you actually call instead of assuming it tracks the first-party price. That is also why Phosphene sends Klein to the Black Forest Labs endpoint first and keeps fal as the fallback: the request is identical either way, and only the bill moves.

What the open weights change

Klein ships open weights under the FLUX Kommunity License, both variants, and the release says they run on consumer GPUs down to an RTX 3090. There are ComfyUI-specific weights and a published workflow.

ComfyUI v0.41.0 landed the same day with native support for the Flux3-image family, so the graph is not a patch job. Two other nodes in that release are worth more attention than their names suggest. ReferenceMultiLatent lets a graph carry several reference latents at once, which is the shape of every consistency problem worth solving: hold a character and a garment and a background in the same stack instead of prompting your way toward an average of them. ImageSDRToHDR converts an SDR still to HDR, which is only useful to you if the rest of the chain can view, store and deliver HDR. Check that before you build on it.

One boundary worth stating plainly: "runs on consumer GPUs" is a compatibility claim, not a throughput claim. The release promises the weights load on a 24GB card. It does not promise they are fast at 4K on one, and those are different sentences.

Local versus API, priced honestly

Once a 4K image costs three and a half cents on the API, "run it locally to save money" mostly stops being true. A thousand 4K images is $36. Add a GPU, electricity, graph maintenance and the hours you spend reading release notes, and the local route needs a different justification: training your own LoRA, building pipelines no API exposes, or work that cannot leave the machine.

That reframing is the useful one. Local buys unlimited attempts at a fixed capital cost. The API buys zero-maintenance attempts at a variable one. At this price level, choose on control, not on arithmetic.

Use the ladder without being fooled by it

Draft cheap, commit at delivery resolution is the standard advice, and it has a hole. Resolution changes what a model can render, so a 1K draft predicts composition, reference roles and lighting well, and predicts fine texture and small lettering poorly. When the thing you are judging is whether the words are spelled right, judge at the resolution you intend to ship.

Bounding boxes are the cheap way around that. They sit in the same request as the resolution value, so you can settle the placement at 768sq and commit the same prompt at 4K. The composition decision happens once, at the price of a thumbnail, and the expensive render is the one you already decided to make.

The release itself is minor news in isolation. A distilled sibling with open weights shows up, some people run it on a 3090, and life continues. What changed is the top of the ladder. If your delivery format is 4K, the question of which model to use was previously answered by your budget rather than by your taste, and it is now worth re-asking. Compare models at the resolution you actually deliver, and the shortlist looks different from the one you built at 1K.