Krea 3 Ultra Enters Arena Image Edit at #4 — $0.06 per 2K Image, 73% Below GPT Image 2
Krea launched Krea 3 on October 15 and by October 19-20 the model had appeared on three independent leaderboards. On Arena’s multi-image edit leaderboard, Krea 3 Ultra debuted at #4 with 1,435 ELO on 2,991 votes, marked preliminary. On the single-image edit Arena leaderboard, it sits at #8 with 1,411 ELO. On Artificial Analysis’s AA-Image-Editing v2.1 benchmark, it places fifth at 1,151 ELO, statistically tied for third with Nano Banana 2.1 and MAI-Image-2.6.
The October 20 update added Ultra thinking mode to the Krea API, FLUX 3 Max video, an Agent API, and Blender integration.
Pricing vs Competitors
The pricing case is sharp. At $0.06 per image at 2K resolution, Krea 3 costs less than a third of GPT Image 2 at $0.22 and less than half of Nano Banana Pro at $0.135. At 100,000 edits per month — a realistic catalog volume for an e-commerce team — the difference is $6,000 for Krea 3 versus $22,000 for GPT Image 2. Krea offers $0.03 at 1K resolution for lower-stakes applications.
| Model | Per-image (2K) | 100K edits/month |
|---|---|---|
| Krea 3 | $0.06 | $6,000 |
| Nano Banana Pro | $0.135 | $13,500 |
| GPT Image 2 | $0.22 | $22,000 |
What Krea 3 Does
Krea 3 is an instruction-following image editing model, not a text-to-image generator (though it supports that as a secondary use case). The model takes an instruction and up to 6 reference images in the standard API, or 8 via the API, and applies the described change while leaving the rest of the image intact.
Three use cases dominate the launch materials. First, product photography: change the setting around a packshot without the label, logo, or product form drifting. Second, character consistency: use reference photos to put the same person in different poses, lighting conditions, or locations across a shoot. Third, precise single-element edits: swap one piece of furniture in a room, change a material color on one surface, or update a text element without touching surrounding content.
The model was built with 2K resolution as the native output size, which Mistral describes as one of the key wins in the training cycle. Models that upscale from lower resolution introduce artifacts at fine text and product labels. Krea 3 avoids the upscale step.
Ultra Thinking Mode
Krea 3 Ultra, the tier that landed on the Arena leaderboards, adds a multi-candidate generation step. The model generates several edits, reviews them against the prompt and reference images, and returns the best result. The process takes several minutes; the API requires a webhook or long-timeout polling pattern. Ultra is available on Max plans and above, costs an additional $0.20 per generation (total: $0.26 per image at 2K), and requires at least one reference image.
FLUX 3 Max Video and Agent API
The October 20 update shipped two features alongside the Ultra API:
FLUX 3 Max video: Generate 5 to 20 second clips with audio at 720p or 1080p, starting from text, an image, or keyframes. FLUX 3 Max is a different model from Krea 3 — video generation rather than image editing.
Agent API and Blender integration: The Krea Agent SDK now runs via HTTP, allowing applications to start agent sessions and send follow-ups programmatically. The same agent can drive Blender scenes and other connected desktop applications, not just Photoshop. A Krea CLI with browser-based sign-in ships alongside it for local development.
Krea 3 in ChatGPT: Krea 3 is available as a selectable model within the ChatGPT image generation UI, giving it distribution through OpenAI’s consumer surface.
Distribution
Krea 3 is available on krea.ai for 30 million users and through fal, Hedra, Together AI, Nous Research, OpenRouter, and Replicate. The breadth of inference partners matters for teams that are already routing image work through one of these providers — Krea 3 can be swapped in without changing the integration layer.
The combination of a strong Arena debut, sub-$0.10 pricing, and day-one availability on established inference platforms puts Krea 3 into direct competition with Nano Banana Pro and MAI-Image for production image editing budgets. Krea 3 is not yet the top-ranked model — GPT-Image-2 and Imagine Image 2.0 hold those positions — but at 73% below GPT Image 2 on a per-edit basis, the ranking gap would need to be substantial to justify the cost difference for most production workloads.