Krea’s own blog dated October 15: Krea 3, an in-house editing model, most edits under 8 seconds end to end at native 2K, $0.06 per image at 2K and $0.03 at 1K. Víctor Perez, co-founder and CEO, says editing models are already in e-commerce, film, and advertising, and that speed was the remaining tax. That is a launch post. It is not a studio booking.
Creative Boom on October 19 is a different clock: a high-quality 2K edit in under six seconds at 3 cents through the Krea API, with Nano Banana Pro at 30 seconds and 15 cents, GPT Image 2.5 at 80 seconds and 22 cents. Startup Fortune on October 16 is a third object: Black Forest Labs’ FLUX.3 Image Klein, open weights, 8GB of VRAM, 5.6 billion parameters, also dated October 15. Three documents. Mix them and you will buy a GPU because a catalog SLA said eight seconds.
We already filed Magnific One at 180 credits for a brand kit and Firefly as a store with Nano Banana as the SKU. This week the edit model’s list price is six cents. The local model’s floor is 8GB.
$0.06 is Krea’s 2K row, not Creative Boom’s 3 cents
Krea’s comparison table, on its blog, is Krea 3 versus Nano Banana Pro versus GPT Image 2. Time per edit: under 8 seconds, 33 seconds, 90 seconds. Price per image: $0.06, $0.135, $0.22. The footnote: Krea 3 at 2K; the other two are median time from request to finished image for edit jobs on Krea, September 24 to October 1, 2026, queue included; list prices without image references.
Creative Boom’s paragraph is not that table. Under six seconds, 3 cents, Nano Banana Pro 30 seconds 15 cents, GPT Image 2.5 80 seconds 22 cents. Different model name on the OpenAI side (2.5 vs 2). Different seconds. Different cents. Do not average 8 and 6. Do not average $0.06 and $0.03. One is the vendor’s 2K list. One is a magazine’s API test. If your finance deck needs a number, pick the row and footnote it. If you pick both, you are lying with a mean.
Volume, still Krea’s blog: 100,000 edits a month at list, no references, 2K. Krea 3 $6,000. Nano Banana Pro $13,500. GPT Image 2 $22,000. That is multiplication, not a discount. Enterprise teams “can talk to us about access.” Talk is not a PO.
Krea 3 is live on krea.ai for what the post calls 30 million users, and through fal, Hedra, Together AI, Nous Research, OpenRouter, and Replicate. fal’s Krea 3 page still says early access, request a form, pay-as-you-go pricing shared with those customers, not printed. Two availability stories. Website users and a 30-million claim on one side. A form on fal on the other. If your pipeline is fal-ai/krea-3, you are in the form. Do not paste $0.06 onto an invoice fal has not sent.
Reference counts split too. Krea’s blog: up to 6 reference images per edit. fal: up to 4 on the edit endpoint, passed as image_urls, referred to in order. If you write a prompt that assumes six faces on fal, you will find the fourth. Give each reference one job, fal says: the person from one image, the jacket from another. That sentence is the actual product. A moodboard dump is how labels smear.
Under 8 seconds is a queue-included median on their jobs
Perez’s quote is about removing the tradeoff between speed and quality at 2K native. The blog also says Krea 3 is not real time yet. Fastest setup: about 4 to 5 seconds. They want that lower. They want more resolution for less compute. Training at 2K was “one of the biggest wins this cycle.” Next is cheaper high-res architectures, plus training data for circling a region or writing on the image instead of prompting. Roadmap is not an SLA. If your art director needs 4 seconds, you are on the fastest setup, not the under-8 median, and you still are not live.
Dominique Yahyavi, co-founder and CEO of Pixelcut, in the same post: they put early checkpoints in front of real Pixelcut users. Lighting looked like a studio. Product stayed the product. Most of their customers are small businesses that cannot afford a studio shoot for every SKU. That is a customer quote on a launch post. Steal the use case, not the adjective “incredible.” The use case is: change the set, keep the label.
Creative Boom’s examples are the same job in magazine language: furniture into an empty room without moving walls, window, raking sunlight; a poster into German in a bus shelter; a billboard reskin; a pencil sketch of a mid-century house into a golden-hour photograph; a marker lamp into a photoreal render. Designers, architects, product people. The through-line is edit-what-I-already-have. Krea’s pitch against generate-from-nothing is the same sentence as Creative Boom’s headline. Believe the constraint. If you do not have a product shot, you are not in Yahyavi’s customer. You are in a text-to-image endpoint that fal also mentions, and that endpoint is not the 8-second edit table.
We already learned Images 2.5’s feature was the edit history. Krea 3’s feature, in its own FAQ, is leaving alone what you did not ask it to touch. History versus restraint. Different buttons. If your complaint last month was identity drift across a catalog, restraint is the row to test. Test it on your SKU, not on Perez’s sentence.
Creative Boom says there are modes from real time to maximum quality. Krea’s blog does not print that ladder in the chunks we have. If you need the ladder, open the app. Do not cite a magazine for a dropdown Krea did not specify here.
8GB is Klein, not Krea 3
Julian Lim, October 16: Black Forest Labs released FLUX.3 Image Klein on October 15, open weights included. Headline number: 8GB of VRAM. 5.6 billion parameters. Floor for mid-range gaming GPUs in bedrooms and small studios. Not a data-center card.
License: FLUX Kommunity License, free commercial use for individuals and companies under $5 million in annual revenue. Cross that line and you need a separate commercial arrangement, or you hit the API. API, via VentureBeat as Lim reports it: Klein starting at $0.009 per 1K-resolution image. Nine-tenths of a cent is not Krea’s six cents at 2K. Different resolution, different vendor, different job. Klein generates. Krea 3’s launch is edit. Do not price a catalog retouch at $0.009 because Klein exists.
History Lim attaches: Black Forest Labs partnered with Krea in July 2025 on FLUX.1 Krea [dev], an open model aimed at the waxy “AI look.” FLUX.2 Klein in January ran inference in under half a second on an RTX 5080 or 5090 at 13GB. FLUX.3 Image Klein in October pushes the footprint down. Partnership in 2025 is not Krea 3 sharing weights with Klein. Do not install Klein and expect Perez’s 8-second 2K edit table.
Krea 3 on a 30-million-user website is a cloud round-trip. Klein on 8GB is a local file. If your legal team will not send product SKUs to krea.ai, Klein is the conversation, and the conversation starts at $5 million revenue and a Kommunity License PDF. If your legal team already lives in OpenRouter, Krea 3’s partner list already named OpenRouter. Pick the constraint first. Then the model.
Lim says BFL claims Klein is still ahead of closed competitors on quality. Claim is not a side-by-side in that piece. Do not put “ahead of GPT Image 2” on a slide from Startup Fortune. Put 8GB and $0.009 at 1K, footnoted to VentureBeat via Lim.
fal is a form. krea.ai is a login
fal’s FAQ is the pipeline version of the launch. Upload the image, add references, describe the edit, name what must stay. One call to fal-ai/krea-3. Send a result back in as the starting image to keep refining. Early access: request the form, install the fal SDK for Python or JavaScript, grab a key. Serverless, no GPUs to manage. Commercial use of images through the fal API is yes, subject to fal’s terms. That last sentence is a license pointer, not a brand-safety review.
How much does it cost on fal? Pricing is shared with early-access customers. Request access. That is not $0.06. If your controller needs a unit cost this week, you are on krea.ai’s printed 2K/1K list, or you are waiting on fal’s email. Do not put “pay as you go, no subscription, no minimum” into a budget as if it were $0.06.
Resolutions and aspect ratios sit behind another fal FAQ that the extract did not fully print. Krea’s blog says native 2K. Creative Boom’s test is a 2K edit. If you need 1K at $0.03, that discount is Krea’s blog, not fal’s form. Keep the invoices in different columns.
Nous Research, Together, Hedra, OpenRouter, Replicate are distribution, not extra quality. Six names on a partner list is how a launch post looks busy. Your legal review is still one vendor at a time. If SKUs cannot leave the VPC, none of the six help, and Klein’s 8GB file is the remaining conversation.
A six-cent edit is not a lighting plot
Launch posts keep teaching the same split: a list price is not a go-live. $6,000 a month at 100,000 edits is what Krea’s multiplication says at list. Your catalog is not 100,000. If it is 2,000 product shots a month, the same arithmetic is $120 at 2K list, before references, before retries, before the fal form. Retries are how $120 becomes a studio day anyway.
Yahyavi’s lighting sentence is the one small businesses will screenshot. Lighting is why you hired a photographer. If Krea 3 keeps the label and fakes the softbox, you still need someone who knows the label was the point. An intern with six references is not that someone if they cannot tell when the nutrition panel warped.
Do not steal Creative Boom’s 3 cents into a Krea SOW that quotes $0.06. Do not steal 8 seconds into a fal queue you have not joined. Do not steal 8GB into a render farm that still bills by the hour.
What you can steal from October 15: edit-first, native 2K, six cents on krea.ai, six references in the blog, four on fal. What you can steal from October 16: Klein is the 8GB open-weight generate path, 5.6B, Kommunity License under $5 million, API at $0.009/1K if you believe VentureBeat through Lim. What you cannot steal is a studio that closed because a CEO said the tradeoff was gone.
Perez wants 4 to 5 seconds to go lower. Until it does, your art ticket still has a photographer column. The flywheel can wait. The SKU cannot.