Vibeleaderboard
Recraft

ai lab

Recraft

Recraft matters as a test of whether a specialist model company can compete through workflow control rather than maximum modality breadth. Its case is strongest where generated work must become editable vectors, layouts, icons, and consistent brand assets, but the scarcity of detailed model papers makes external technical assessment harder than product assessment.3,5,10

Official site ↗Models in the Atlas →

Profile

Overview

Origins in machine learning and design

Recraft is an image-generation and editing company founded in the United States in 2022. Recraft currently describes its headquarters as London, while independent reporting previously described the company as San Francisco-based. Its founder and chief executive, Anna Veronika Dorogush, previously co-authored CatBoost, an open-source gradient-boosting library. Recraft was built around a narrower problem than a general image chatbot: helping designers generate and revise usable visual assets inside a production workflow.1,2,9

A model inside a design workspace

The product combines proprietary image models with an infinite-canvas editor, collaboration, an API, and tools for inpainting, outpainting, mockups, background removal, upscaling, and vectorization. Its model and API taxonomy distinguishes raster styles from vector and icon styles, with vector outputs delivered as SVG rather than as traced raster previews. This makes format, layout, and editability part of the product definition rather than post-processing left to another application.1,3

V3 and the red_panda reveal

Recraft V3 was released in October 2024 with controls for text size and position, style, inpainting, and outpainting. Before the company disclosed its identity, the model appeared as red_panda in a blind image arena and drew attention for its preference score. Independent reporting later connected that benchmark result to Recraft and reported that the company had reached four million users and more than $5 million in annual recurring revenue when it announced a $30 million Series B in May 2025.4,9

V4 and the production-design thesis

Recraft V4 arrived in February 2026 as a ground-up model rebuild offered in standard, high-resolution, vector, and high-resolution vector variants. Recraft followed it with V4.1, whose Utility Pro variant entered the Artificial Analysis Image Arena in May and was ranked twenty-third on the live text-to-image leaderboard when this profile was reviewed. The company describes the family as tuned with designers for composition, color, typography, and production assets. One independent product review found Recraft unusually useful for vectors, mockups, repeatable styles, and brand work. Recraft's current terms confirm that credit use, asset ownership, and training permissions vary by plan and require buyer attention.5,6,7,10,8

Notable contributions

  1. 01Prompt-to-SVG generationRecraft integrated native vector and icon generation with SVG delivery, allowing generated assets to remain scalable and editable in downstream design tools.3,5
  2. 02Explicit text-layout controlsV3 added controls for text size and placement alongside improved text rendering, inpainting, and outpainting, addressing layout tasks that matter in posters, packaging, and campaign graphics.4
  3. 03Brand-style consistencyThe product treats reusable visual styles and consistent brand output as first-class controls available in the editor and API rather than as prompt-only conventions.1,3
  4. 04Blind testing before model disclosureRecraft tested V3 under the red_panda alias before identifying the model, giving the launch an externally observed preference result instead of relying only on a named vendor comparison.9