Krea 3 Open Beta: A Frontier Image Editing Model Claiming 4× the Speed of Nano Banana Pro

Krea 3 is a new frontier image editing model from Krea, launched in open beta, that is positioned around production economics rather than raw generation: it claims to be roughly 4× faster than Nano Banana Pro and to cost about $0.06 per 2K edit versus roughly $0.135 for Nano Banana Pro and roughly $0.21 for GPT Image 2.5. The announcement reached the front page of Hacker News with 270 points and 86 comments, where the debate centered less on whether the model is fast and cheap and more on whether its output quality, pricing flow, and closed weights make it usable.

What Krea 3 is

Krea 3 is an image editing model, not a text-to-image generator. Krea describes it as "an image editing model built for production" that "changes only what you ask." The company's own framing on the landing page is that it is a frontier-quality model at a fraction of the speed and price of competing frontier editors.

Krea's co-founder, posting on Hacker News as dvrp, described the release directly:

"Today we are announcing the open beta of Krea 3, our newest foundational image model. In a nutshell, it's a frontier image model in terms of quality, but at a fraction of the speed and a fraction of the price."

The same comment states that a more technical write-up is planned, following the pattern of Krea 2's technical report, and that the launch page focuses on capabilities rather than benchmarks.

The headline numbers: speed and price

Krea's comparison block runs one image through twelve identical prompts across three models and reports median request-to-result times including queue time:

• Krea 3: under 8 seconds per edit, approximately $0.06 per edit
• Nano Banana Pro: approximately 26.8 seconds per edit, approximately $0.135 per edit
• GPT Image 2.5 (Sunburst): approximately 39.9 seconds per edit, approximately $0.21 per edit

Krea labels these as "pre-generated outputs, not a live benchmark," and states that GPT Image 2.5 pricing is "a draft estimate pending validation." The landing page also advertises "4× faster than Nano Banana Pro" and a "10× Faster, 3× Cheaper" headline. The numeric price and speed chart elsewhere on the page rendered as placeholder values in the scraped source, so those specific per-model figures cannot be verified from the available text.

Editing capabilities Krea 3 advertises

Native regional edits. Draw a box and describe the change; Krea 3 redraws inside the box and leaves the rest of the image as it was. Krea says changed pixels are measured against the original.

Hex-code color matching. A prompt can name a hex code, such as #FFC12E, and Krea 3 matches that color on the product and only on the product. The page demonstrates this across plush toys, a 911, a desk fan, a dress, a kitchen island, a glass stand, lip gloss, a sofa, and a tunnel.

Up to eight references in one image. Users can point at people, products, places, and styles by number, and Krea 3 composites them into a single picture.

Character consistency across edits. Krea 3 can change styling, setting, and mood while keeping a face recognizable across a sequence of edits.

Native 2K output. Krea 3 renders 2K natively; the sample image on the page is 2736 × 1536 px.

Voice-note prompting. Krea describes sending a photo followed by a voice note, with the edit returned before the user finishes their next thought.

Layout-preserving design edits. Krea shows poster, packaging, editorial, and advertising examples where the product, photo, or brand is swapped while layout and type are held.

Precision claims versus competitors

Krea publishes a pixel-difference comparison for the prompt "Change only her coat to cobalt blue," showing heat maps of changed pixels for Krea 3, Nano Banana Pro, and GPT Image 2.5. The company defines precision as "the share of pixels outside the edit area left unchanged," measured on one attempt per model at 2K with the same image and prompt, with Krea 3 running in Precise mode. Outputs are compared at quarter resolution after best alignment so that a different output size does not count as a change.

What Krea says about benchmarks and future plans

Krea acknowledges that public benchmarks are not ready. In the launch comment, dvrp wrote that "benchmarks still being worked on, but in our internal evals it seems a no-brainer replacement to NBP or GPT Image, both speed and cost-wise."

On roadmap items, the same comment states that the team is considering letting users fine-tune Krea 3 via LoRAs and is considering open weights, but that those decisions come after launch issues are resolved. No compute or training-time figures were disclosed; a Hacker News user asking for "any vague details on how much compute/train time was needed" received no answer in the thread.

What Hacker News commenters found strong

Commenters broadly agreed that speed and cost are the differentiating axes. One user building a personal styling app wrote that fast, affordable image editing "unlocks tons of cool use cases" and that "it's refreshing to see latency be held just as important as price as-well."

A detailed UI-design comparison against GPT-Image-2.5 found Krea 3 weaker on visual creativity but notably better at one specific thing:

"Overall Krea 3 isn't near GPT-Image-2.5 in terms of visual creativity in this space, but it still should be commended for its ability to render small text details."

The same commenter added that they would "love to get my hands on this to post-train and see if it can be pushed further in this space," and said they "actually see a lot of potential for the model."

What Hacker News commenters criticized

Output quality gaps. Several users reported failures on specific edits. One said "the pearl earring and the billboard worker didn't feel like frontier quality." Another reported that asking Krea 3 to seat a woman in a T-shirt at a fancy bar caused the model to invent a suit jacket, with the shirt's graphic running over the jacket, and that "Gemini did much better." Another noted scrambled signage text in the fuzzy-characters scene, calling it "a no-go." A further comment flagged an unwanted shadow added to a red car.

Hex codes leaking into output. Despite the hex-matching feature, one commenter reported that specified hex codes "tend to leak into the design," pointing to a hex code appearing in the valuation panel of a generated banking website.

Pricing flow and dark patterns. Multiple commenters objected to the signup experience. One wrote that Krea "offers you plans when you sign up, if you decline them, it offers you a discount," adding: "Just offer the discounted price up front. If your business cannot survive without sneaky pricing, it's not a good business." Another called the flow "incredibly aggressive dark patterns in this sign up flow just to try the model." A third objected to the trial structure: "2 images to try and then the only option after is a monthly subscription. I don't mind buying credits but I don't want a subscription."

Cost still too high for some product use cases. One commenter argued Krea 3 is "still to expensive for most 'product' use cases," and noted that their own best results-per-dollar still came from Grok Imagine V1, "a now deprecated model."

Missing comparisons. One user asked why there is no comparison against FLUX 3 Image. Another asked whether Krea 3 is related to Krita; a reply clarified the scope with "For image editing."

The open-weights question dominated the thread

The most repeated request in the comments was for open weights. One commenter asked directly whether open weights are being considered, citing interest in anime fine-tunes. Another wrote: "Meh ... there are tons of great paid image models ... where the open weights at!?"

A longer comment framed the strategic tradeoff:

"Open-weights might make this wildly popular but I could see this not making headlines and grabbing attention if it remains closed since most people will just continue using OpenAI and Google. I know an open-weights release is being considered so fingers crossed. Hoping to run this locally for my personal use cases since I don't trust any hosted providers with personal image data."

Krea's own comment confirms open weights and LoRA fine-tuning are under consideration but not committed.

Caveats on the available data

The performance and pricing figures in this article come from Krea's own marketing page and its own comparison methodology, which Krea explicitly labels as pre-generated rather than a live benchmark. Independent benchmarks for Krea 3 were not available at the time of the launch thread, and Krea stated that its internal evaluations were still in progress. The scraped landing page contained placeholder values in its price and speed chart, so only the figures stated in the explicit comparison block are reported here.