Startup Krea has opened the beta for Krea 3, a new image editing model. According to the company's own measurements, editing in native 2K resolution takes less than 8 seconds, including queue, and costs $0.06 per operation. The beta is already available even on the free plan, and developers have access to the krea/v3/edit API endpoint on the fal platform.

image
image
image

What happened

The Krea announcement is dated October 15, 2026. The beta is open on Pro, Max, Business plans and on the free Free plan with a limit of about five images per day. The price of a single edit is stated at $0.03 per 1K image and $0.06 for 2K; for comparison, Krea cites competitors' price lists: Nano Banana Pro costs $0.135 and takes 33 seconds to edit, Nano Banana 2 fits into 27 seconds, and GPT Image 2 costs $0.22 and 90 seconds. The model accepts up to six references per operation, preserves products, characters, and text between edits, and supports area selection via bounding box. For developers, the model is hosted on fal under the krea/v3/edit endpoint with pricing of $0.025 per megapixel, where one megapixel corresponds to a resolution of 1024×1024; the interface offers a choice between 1K and 2K and a fixed seed. In addition to fal, Krea 3 is available through Hedra, Together AI, Nous Research, OpenRouter, and Replicate, and in Weavy AI it is connected via Import Model.

Context

Krea 3 belongs to the class of edit models: such systems do not draw an image from scratch, but make pinpoint changes to an existing one, which is critical for product catalogs, advertising creatives, and localized posters. Until now, according to Krea's data, production editing was limited by price and speed: editing remained an expensive and slow manual iteration, not suited for mass volume. The stated response is positioned as an engineering and economic step, not a scientific breakthrough: the model starts immediately with a production package — an open API, public pricing, and latency instead of closed demos. The statement about speed is disclosed in more detail than usual marketing: the median edit in 2K taking the queue into account, the measurement window from September 24 to October 1, 2026, and the owner of the metric, which is Krea itself, are specified. Quality is positioned at the level of Nano Banana Pro and GPT Image 2.5.

Why this matters for the industry

For the industry, the bet is on the economics of mass pipelines. According to calculations from the announcement, 100,000 edits per month at price lists will cost about $6,000 versus $13,500 for Nano Banana Pro and $22,000 for GPT Image 2, meaning that at large volumes the savings exceed two times. With short latency and a fixed seed, pinpoint editing ceases to be a one-off manual operation: it can be called from pipelines for processing catalogs, creatives, and poster localization as a normal reproducible function. The announcement creates price pressure on the entire segment of edit models: if Krea 3's price list is confirmed on real loads, previous edit rates are unlikely to be maintained. Further, the roadmap expands the effect: cheaper high resolution and new interaction modes, such as outlining objects directly on the image, can make cheap editing a standard step in production pipelines and agent frameworks. If the trend in price and speed is maintained, in the long term value will shift from the editing operation itself to the consistency of brand assets and reproducibility of results.

Why this matters for users

You can try Krea 3 right away: the beta is open on the free Free plan, so the first trials do not require payment, and Pro, Max, and Business plans are provided for regular work. Developers have access to the krea/v3/edit sandbox on fal with a choice of 1K or 2K resolution and a fixed seed: with the same input data, the result is repeated, which is convenient for comparing edit variants. A practical tip from the announcement — start the prompt with what should remain unchanged, for example "keep the product exactly as it is", and specify what is on which of the input images: according to Krea's data, this noticeably increases the accuracy of pinpoint edits. A reasonable first step is to run your own set of tasks and compare latency and quality with current tools, without moving production immediately.

What is still unknown / limitations

The main limitation is that all key figures belong to Krea itself: the median of less than eight seconds is formulated more specifically than usual marketing claims, but the methodology is not disclosed — the prompt sample, hardware context, and time distribution in tails like p95 and p99 are unknown. The stated quality at the level of Nano Banana Pro and GPT Image 2.5 is a self-assessment without public benchmarks, user ratings, and comparison methodology. The described capabilities — preserving products, characters, and text between edits, accepting up to six references, and area selection via bounding box — are promises, not measured properties, and the key undisclosed question is how quickly artifacts accumulate in long chains of sequential edits. Until independent measurements, the figures should be treated as a hypothesis, although it can be checked directly via the API on your own set of tasks.

Sources

Author

Look at AI, editorial team