Krea Launches Krea 3 to Edit 2K Images 71% Cheaper Than GPT

Krea has opened beta access to Krea 3, its proprietary foundation model for editing existing images. The company reports end-to-end 2K edits in under eight seconds, with prices below competing models from Google and OpenAI. Krea also says it is the only independent lab represented in the top 10 of the public image-editing arenas it tracks.

Fast and Ultra split the latency budget

Krea 3 offers two inference modes. Fast prioritizes latency and is responsible for the advertised sub-eight-second 2K edits. Ultra allocates more compute to image fidelity and ranked fourth across the arenas Krea evaluated, behind three OpenAI image models.

On Alibaba’s CPI-Bench, Krea reports a score of 4.81 out of 5, placing Krea 3 second behind GPT Image 2.5. Leaderboard positions can change as models and votes are added, and the reported results should be treated as a snapshot rather than a fixed ranking.

Six cents for a 2K edit

Krea lists beta pricing at $0.03 for a 1K-resolution edit and $0.06 for a 2K-resolution edit. Based on the company’s comparison prices, the 2K rate is about 56% lower than Nano Banana Pro and 71% lower than GPT Image 2.5.

Krea attributes the lower latency and cost to regional computation. The model concentrates processing on the selected area while preserving pixels elsewhere, reducing the work required for localized changes. Krea has not published enough architectural detail to independently assess that explanation.

Regional edits keep the rest intact

Krea 3 is designed for controlled changes to an existing image. A user can draw a bounding box, describe the modification, and continue refining the result over multiple turns while the model attempts to preserve faces, lighting, composition, and backgrounds outside the selected area.

• Bounding-box editing: Restricts a requested change to a selected region.
• Reference images: Supports six to eight references, depending on the workflow, according to Krea.
• Multi-turn editing: Applies successive instructions while retaining prior changes.
• Hex color prompts: Accepts hexadecimal color values for named objects.
• Native 2K output: Produces 2K results directly instead of relying on a separate upscaling step.

Regional processing is useful when a workflow requires one controlled alteration, such as changing a jacket’s material or recoloring product packaging, without redrawing the full frame on every turn.

Company benchmarks favor Krea on speed and price

The table uses Krea’s measurements and published rates. Actual latency will depend on provider overhead, queueing, region, concurrency, input size, and selected inference mode. Krea has not disclosed a complete reproducible methodology covering identical hardware and service conditions for every model.

Hosted APIs, creative plugins, and licensed weights

Hosted access is available through Krea and inference providers including fal, Together AI, Baseten, and other partners. Provider-specific limits, SDKs, authentication methods, rate limits, and billing terms vary.

Krea also lists integrations for Photoshop, After Effects, Premiere Pro, DaVinci Resolve, Blender, and ChatGPT. These integrations place regional editing inside established design, video, and 3D workflows.

Organizations can request a commercial license for the model weights to self-host inference, fine-tune on proprietary assets, or meet data-residency requirements. The weights are not openly released, and Krea has not published standard licensing prices or hardware requirements.

Production editing is the clearest fit

Krea 3 suits workflows that preserve most of an image while changing a specific element. Likely applications include e-commerce variants, packaging mockups, campaign adaptations, film-plate relighting, localized corrections, and character edits across a sequence.

Krea 3 does not generate images from a blank text prompt. Krea 2 remains the company’s text-to-image model and has open weights, while Krea 3 is a proprietary editing model available through hosted services or a commercial weights license.

Krea reports that its fastest configuration can return edits in roughly four to five seconds, although the broader launch claim is under eight seconds end to end. The company does not describe the current model as real-time, which limits its use for interactive video or frame-by-frame live applications.

Fine-tuning remains on the roadmap

Krea lists LoRA training and broader style-reference support as planned additions rather than current beta features. LoRA is a lightweight fine-tuning method that adapts a model with a relatively small set of additional parameters, which could help teams reproduce products, people, and visual styles more consistently.

Production evaluation still requires details that Krea’s launch material does not provide, including service-level commitments, throughput under concurrency, content-policy limits, retention terms, supported input formats, and commercial licensing conditions. Those factors will determine whether the benchmark advantages carry into deployed systems.