AlphaMaven AlphaMaven

Krea 3 Hits Benchmark Second Place and Ships the First Trainable Image Editor

Reported by Glitchwire
Oct 15, 2026· 3 min read
Save Post LinkedIn Email
Key Takeaways
  • • Krea released Krea 3, an image editing model for high-volume commercial work, and its Ultra tier scores 4.81 out of 5 on CPI-Bench, an Alibaba Group benchmark for real-world editing, placing it second behind OpenAI's GPT Image 2.5 according to Glitchwire.
  • • Krea claims an edit at 2K resolution completes in under 8 seconds end-to-end, roughly 10 times faster than GPT Image 2 and 4 times faster than Nano Banana Pro, priced at $0.03 per 1K image and $0.06 per 2K image.
  • • Krea 3 is marketed as the first edit model that can be LoRA-trained on a specific product or person, allowing teams to keep characters, products, or visual styles consistent across campaigns.
  • • Krea, which says one in three Fortune 500 companies use its platform and has raised over $80 million from investors including Bain Capital Ventures and a16z, is distributing the model via krea.ai, its API, and partners including fal, Hedra, Together AI, Nous Research, OpenRouter, Baseten, and Replicate.
Summary
Krea, a San Francisco company that reports a 35-million-user creative platform built around its own image models, released Krea 3, an image editing model aimed at high-volume commercial work. On CPI-Bench, an Alibaba Group benchmark for real-world editing tasks, the Ultra tier scores 4.81 out of 5, ranking second only to OpenAI's GPT Image 2.5 and ahead of GPT Image 2, Seedream 5 Pro, and Nano Banana, according to the source article.

Speed is the central differentiator Krea is emphasizing. Krea says an edit at 2K resolution completes in under 8 seconds end-to-end, which it characterizes as 10 times faster than GPT Image 2 and 4 times faster than Nano Banana Pro. Pricing is listed at $0.03 per 1K image and $0.06 per 2K image. The model also accepts up to six reference images per edit, supports native 2K output with 4K for print work, and offers bounding boxes for precise element placement.

The larger shift, in Krea's framing, is LoRA fine-tuning, which it describes as "the first edit model you can train on a specific product or person." The company argues this allows teams to hold a character, product, or style constant across a campaign, a capability relevant to e-commerce catalogs and advertising series. Co-founder and CEO Víctor Perez said editing models have been "held back by speed" and that faster edits let creatives "stop rationing their ideas."

Early users include Pixelcut, an AI product photography platform used by more than 70 million creators, which has tested Krea 3 since early checkpoints. Pixelcut co-founder and CEO Dominique Yahyavi said most of its customers are small businesses that cannot afford studio shoots. Doodles, the Web3 entertainment company, is also among initial users. Krea says one in three Fortune 500 companies now use its platform and has raised over $80 million from investors including Bain Capital Ventures and a16z. Justine Moore, an a16z partner who helped lead the investment, said fast edits turn the model "into the workflow."

Krea 3 is available on krea.ai, through the Krea API, and via distribution partners including fal, Hedra, Together AI, Nous Research, OpenRouter, Baseten, and Replicate. Enterprise customers can request access through Krea Partnerships. The source notes that whether the benchmark results translate to production pipelines remains to be seen.

Go deeper than the headline

Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.

Start your free 7-day trial

Sources

1 outlet

AlphaMaven summarised this story from the reporting below. Each link opens the original article.

Glitchwire Krea 3 Hits Benchmark Second Place and Ships the First Trainable Image Editor