Recraft: The Red Panda That Learned Design Taste

A Panda Nobody Could Explain

In the last week of October 2024, the small tribe of people who refresh AI leaderboards for fun noticed something strange. A model nobody had heard of had climbed to the very top of the Artificial Analysis text-to-image arena. It had no company name attached, no launch post, no press release. It went by a single lowercase codename, red panda, written with an underscore in the middle the way engineers name things when they do not expect anyone outside the building to see them.

The arena works a bit like a blind wine tasting. A visitor types a prompt, two anonymous models each draw a picture, and the visitor picks the one that better matches what they asked for. Thousands of those votes get turned into an Elo rating, the same ranking system chess uses to compare players. And red panda was not scraping by. It sat roughly 40 Elo points ahead of the previous leader, Flux 1.1 Pro from Black Forest Labs, which at the time was the model serious people were excited about. It was also fast, generating an image in a median of around 7 seconds, roughly twice as quick as OpenAI's DALL-E 3.

TechCrunch ran a story about a mysterious new image model. People on X posted the images it made and argued about who could be behind it. Was it Google testing something in disguise? Was it OpenAI? A Chinese lab? Mystery models had become a marketing genre by then, a way to build anticipation before a reveal, so everyone assumed a big name was hiding behind the panda.

It was not a big name. A few days later the model was revealed as Recraft version 3, made by a startup that most people in the field knew, if they knew it at all, as the company that generated vector icons. A team that had raised a modest Series A less than a year earlier had just beaten Midjourney, OpenAI, and Black Forest Labs on the most watched public scoreboard in image generation. And the codename? According to founder Anna Veronika Dorogush, it stuck because early testers just kept generating pictures of red pandas.

That is the hook. But the more useful story, especially if you are trying to decide which image model to open on a Tuesday morning, is why a company built around icons and logos ended up there at all, and what it has turned into since. Because Recraft today is a different animal from the panda of 2024, and knowing what it is actually good at will save you money and frustration.

The Problem With Pretty Pictures

Go back to 2022 and early 2023. Midjourney, DALL-E, and Stable Diffusion were producing images that made people gasp. Astronauts on horses. Paintings in the style of anyone. Product photos of products that did not exist. If you were a casual user, it felt like magic. If you were a working graphic designer, it felt like magic that did not fit through the door of your office.

Here is the first problem. Every one of those models painted pixels. A pixel image, what designers call a raster image, is a mosaic of tiny colored tiles. It looks great at the size it was made, and it falls apart when you blow it up. Now think about what designers actually make all day. Logos that need to sit on a business card and on the side of a truck. Icons that need to be crisp at 16 pixels and at 512. Illustrations that go to a print shop. Those things live as vectors.

A vector image is not a mosaic at all. It is a recipe. Draw a curve from this point to that point, bend it this much, fill the shape with this exact green. Because it is a set of instructions rather than a grid of dots, a computer can redraw it at any size, and it stays perfectly sharp. The standard file format is SVG, short for scalable vector graphics. And crucially, a designer can open that file and grab one curve, move one point, change one color, without touching anything else.

So a generator that only paints pixels is, for a lot of design work, like a print shop that only delivers photographs of your business card. Pretty, but not the thing you needed. You could trace the raster image into a vector afterwards, and plenty of tools would try, but tracing is a lossy chore. You get wobbly edges, too many points, shapes split in odd places. Often it is faster to redraw from scratch.

The second problem was text. Early image models treated letters as decorative squiggles. Ask for a poster that said Grand Opening and you might get Grnad Opeinng, or something that looked like a language from a dream. The third problem was brand control. A company has specific colors, a specific mood, a specific level of detail. The models ignored all of that unless you got lucky.

And the fourth problem, the sneaky one, was consistency. Generating one beautiful image is easy. Generating the eighth image in a campaign, so that it clearly belongs with the first seven, is hard. Without some way to lock a style, every image was a talented stranger doing its own thing. For a single social post, fine. For an icon set of forty symbols, useless.

That gap, between images that impress and images that ship, is the gap one machine learning engineer decided to build a company in.

The Woman Who Built CatBoost

Anna Veronika Dorogush did not arrive in image generation from art school. She studied mathematics and computer science at one of Russia's top universities, and while she was doing that, she worked as a professional fashion model. Those two worlds rarely share a person. One is about precision and proof, the other about how an image lands on a viewer in a fraction of a second. In hindsight it is hard not to read that combination as the whole thesis of her later company.

Her engineering career ran through some of the biggest names in technology. She had stints at Microsoft and Google, and then spent years at Yandex, the Russian search giant, where she rose to head of machine learning systems. At Yandex she co-created the thing she is best known for in the data science world, an open source library called CatBoost.

CatBoost is not glamorous, and that is exactly the point. It is a gradient boosting library, which means it builds a large crowd of small decision trees, each one correcting the mistakes of the ones before it, until the crowd makes good predictions. Its particular trick is handling categorical data well. That is data like city names, product types, or the brand of someone's phone, labels rather than numbers, which many algorithms stumble on. CatBoost became a workhorse for tabular data, the spreadsheets and databases that most real businesses actually run on. It is the kind of tool that quietly sits under recommendation systems and forecasting pipelines and almost never makes headlines.

So here is someone whose reputation was built on the reliable engine under the hood, not the shiny demo. When she later talked about her modeling years, the lesson she drew was about focus rather than effort.

The biggest lesson from that time was that grinding isn't everything.

She went on to explain that when building a company, you have to be excellent at the things that are mission critical, and not spread yourself across everything else. Keep that in mind, because it explains a lot of Recraft's choices. It did not try to be the best at everything. It picked a lane where it thought the giants were weak.

She founded Recraft in 2022. It is also notable, in a field crowded with co-founder teams of former big lab researchers, for having a solo female founder and chief executive. The company has never made much of its name, but it reads like a statement of method. Not craft from nothing, but craft, and then craft again, the iteration that every designer recognizes.

Out of Stealth and Into the Model Business

Recraft came out of stealth on May 31, 2023, with a launch on Product Hunt. The headline feature was the one nobody else had, generating vector graphics from a text prompt. Not tracing, not converting, generating. Eight months later the company reported more than 300 thousand users, which for a niche professional tool was a strong start.

In January 2024 it announced a Series A of twelve million dollars, led by Khosla Ventures, with former GitHub chief executive Nat Friedman also in the round. At that point the team was thirteen people, based in London. And Dorogush used the money to make a decision that sounds obvious now and was bold then. She would build her own foundation model from scratch rather than fine tuning someone else's.

Her reasoning, as she explained it to TechCrunch at the time, went like this. You can take an existing open model and teach it your style, your brand colors, your preferences. But if the base model fundamentally cannot draw a pool player or a ballet dancer properly, no amount of fine tuning on top will fix that. The foundation has to be right. If you want real control over style, consistent output, brand colors and level of detail, you need to own the model all the way down.

The investors bought the argument. Nikita Shamgunov at Khosla Ventures framed it as a gap in the market. Most generative design tools, he said, were aimed at consumers, while professionals need a high degree of control, things like vector output, style controls and full production workflows, all powered by a model the company built itself.

In March 2024 came Recraft version 2, the company's first model trained entirely from scratch. It had roughly 20 billion parameters, and it was trained on infrastructure from Nebius, the cloud company that grew out of Yandex's international business, which is a nice little loop for a founder who came from Yandex. Version 2 brought native vector generation into the model itself, plus conditioning on brand colors and noticeably more consistent icons and illustrations. It is still on sale today as the cheapest option in the lineup, about two cents per raster image, and Recraft still recommends it for minimal icon work.

Then, in October 2024, the panda. Version 3 took everything up a level. Much better photorealism, legible text across multiple words, closer obedience to long descriptive prompts, and a new artistic dial that let you turn stylistic intensity up or down. That was the model that walked onto the leaderboard anonymously and took first place. For a company of that size, it was the equivalent of an unsigned band headlining a festival.

Money, Users, and the Hub Strategy

Success on a public leaderboard does wonderful things to a fundraising deck. In May 2025, Recraft announced a thirty million dollar Series B led by Accel, with Madrona joining and Khosla returning, bringing total funding to about forty-two million dollars. At the announcement the company reported more than 4 million registered users. Later reports put the figure above 7 million, spread across more than 200 countries. A startup tracker also notes that NVIDIA joined an extension of that round in the summer of 2025.

Forty-two million dollars sounds like a lot until you put it next to the companies Recraft was now competing with. OpenAI and Google spend that kind of money on compute without blinking. And over 2025, that difference began to show. Google's Nano Banana models and OpenAI's GPT Image models leapt forward, especially at editing existing images and following complex instructions. The red panda's crown did not last forever. By May 2026, according to one industry tracker, every model ranked above Recraft's best on the main leaderboard belonged to either OpenAI or Google.

So Recraft did something that looks counterintuitive at first. It opened its doors to the competition. Recraft Studio, the company's web workspace, an infinite canvas where you generate, edit and arrange images, started hosting rival models right alongside its own. Today, inside Studio, you can pick Nano Banana, GPT Image, Flux, Ideogram, Imagen, Seedream, Qwen Image, and more. It even added video models like Veo, Kling, Sora and Seedance.

Why would a model company sell other people's models? Because it turns Recraft from one contender in a horse race into the stable where designers keep all the horses. If Google's model wins this month and OpenAI's wins next month, the designer still opens Recraft, still works on Recraft's canvas, still uses Recraft's vectorizer, upscaler, background remover and mockup tools. The bet is that the workspace and the design specific features are the durable business, and the model leaderboard is weather.

Two more moves extended that reach. In July 2025, Recraft released a server for MCP, the Model Context Protocol, the standard that lets AI assistants call outside tools. That meant you could ask Claude or a coding tool like Cursor to generate a Recraft image directly, without opening the app. And in March 2026, Picsart, the mobile design platform with more than 130 million creators, announced an exclusive partnership to give its users early access to a new Recraft feature called Exploration Mode. That is a huge distribution channel, and also a dependency. If a partner that size changes priorities, a smaller company feels it.

Version 4 and the Bet on Taste

On February 17, 2026, Recraft released version 4, and with it a new argument about what an image model should be for. The company described it as a ground-up rebuild, developed in close collaboration with designers. And the word it kept returning to was taste.

The pitch went like this. The leading image models, Recraft argued, are tuned for broad general preference. They are trying to make an image that the most people will like, which is a reasonable goal for a consumer chatbot. But mass appeal is not the standard for a brand system or a campaign. Designers want composition with a clear hierarchy, color relationships that feel deliberate, a place where the eye goes first. Recraft said version 4 was tuned to make those aesthetic decisions, not just to tick off the objects in the prompt.

The launch post was full of side by side comparisons, mostly against Google's Nano Banana Pro. One was a poster with the word Overthink built as inflated, glossy green 3D letters inside a deep shadow box frame. The prompt asked for the letters to look compressed, pressing against the edges of the frame. Recraft's point was that its model actually made the letters push against the box and create tension, while the competitor drew nice inflated letters floating politely in the middle. Another prompt asked for exactly three sage green bottles and one matching jar for a fictional brand called Botanica Lab. Recraft claimed its model got the count and the labels right, and that the rival produced the wrong number of products and missed branding on one bottle.

Version 4 came in a family. A standard version for fast, cheap iteration, about 10 seconds per image at one megapixel. A Pro version for print and large formats, about 28 seconds at four megapixels. And vector variants of each, so the same aesthetic could come out as editable SVG files. Recraft made a point of opening all four to free users too.

A month later, in March, came Exploration Mode, the feature Picsart got early. Instead of writing an ever longer prompt and hoping, you give one idea and the system generates a spread of different visual directions. You pick the one that feels right and refine from there. It is a quiet admission of something designers have always known. Often you do not know what you want until you see a few things you do not want.

Then on May 14, 2026, version 4.1. This is the current flagship, and it split the model into three personalities, which is the single most useful thing to understand if you plan to use it.

The main version 4.1 model is the expressive one. Recraft describes it as bringing its own point of view to every prompt, experimenting with light, mood and composition like a creative director would. It is especially good at short prompts. In one of Recraft's comparisons, the prompt was just five words, a wristwatch, metal and glass, close-up photo. Recraft's model produced a dramatic editorial shot of brushed metal catching light. The rival model, in Recraft's telling, put the watch on someone's wrist, on a park bench. Technically correct, aesthetically elsewhere.

The second personality is version 4.1 Vector, built for logos, lettering and illustration, where the quality lives in the line. The third is version 4.1 Utility, and this one is easy to overlook and very handy. Utility deliberately turns off the drama. Flat lighting, front facing composition, simple scenes. It is meant for mockups, product shots, catalog images, anything where a calm and predictable picture serves the work better than an artistic one. A tracker reported that at the end of May, the Utility Pro variant ranked third worldwide on the Artificial Analysis image arena.

A fair warning about all of these comparisons. They come from Recraft's own blog. Every company picks the prompts where it wins, and some of the commentary is gleefully catty, including a jab that OpenAI's model defaults to a slightly cartoonish face. Treat the examples as a description of what Recraft is aiming for, not as an independent verdict.

A Summer of Shipping

If version 4.1 was the engine upgrade, the months after it were about the workshop around the engine. And the pace was relentless.

On July 23, 2026, Recraft added direct vector editing inside Studio. Double click any shape in a generated vector and you can move anchor points, bend curves, recolor individual shapes, and duplicate elements, all on the canvas, without exporting to Illustrator or Figma first. That sounds small. It is not. It means a generated logo stops being a take it or leave it object and becomes a first draft you can actually fix.

Six days later came the Design Agent. You describe your product in plain language, and the agent builds a consistent set of brand assets from it. A logo, a color palette, typography, icons and more, all saved together as a reusable design kit, so later generations stay in the same visual world. For a small business that has never had a brand guide, that is close to a starter identity in one sitting.

August brought a redesigned Studio interface, with a create panel and a chat panel for the agentic, conversational way of working. Select any image and you see the exact prompt, model and settings that made it, with buttons to modify it in place, reuse that recipe for something new, or finalize it with upscaling and export. Recraft added batch generation of up to 8 images at once, a compare models switch that runs two engines side by side on the same prompt, and voice input, so you can describe what you want out loud.

The most important August release, though, was version 4 Styles. Pick a style from a library, or build one from your own reference images, and every image you generate afterwards holds to it, down to the finer details, with no training step. This goes straight at problem number four from the beginning of this episode, the eighth image drifting away from the first seven. Through the application programming interface, styled generation starts at about three and a half cents per image.

September added OpenAI's newest image model to the Studio menu, and then, on September 23, version 4.1 Flash. Recraft claims it generates an image in 1.3 seconds end to end, at a fraction of the normal cost. Paired with it is a new Refine button that takes a quick draft you like and lifts it to high definition. Draft at the speed of thought, then pay for quality only on the keepers. That is a workflow, not just a model.

The Case Against

So far this has sounded like a highlight reel, and it would be unfair to leave it there. Recraft has real critics, and some of their points matter for anyone deciding whether to spend money on it.

The design publication DesignWhine has reviewed Recraft repeatedly, and its current score is 3.7 out of 5, which on its scale means good but with weaknesses that affect the recommendation. On the positive side, the reviewers called native vector generation and cross-image style consistency genuine innovations that address real professional needs. On the negative side, they reported reliability problems in their testing, inconsistent dimension controls, projects that failed to load, and AI responses that were hard to predict. They also called out a steep learning curve.

Their sharpest critique is about money. Recraft runs on credits. Different models and operations cost different amounts, and conversations with the agent can consume credits on top of the images themselves. That makes it much harder to budget than ordinary design software with a flat seat price. A designer who explores a lot can burn through a plan very differently from one who makes a handful of final assets. There is also a rights distinction worth knowing. Recraft says paid plans grant ownership and commercial rights for images made while subscribed, and the free plan's terms are more restrictive. If the work is for a client, that is not small print.

Then there is the taste argument itself. The reviewers made a point that is easy to nod past. Taste is contextual. A lot of brand work requires restraint, and a model with its own point of view can fight you when what you need is plain and quiet. That is presumably why the Utility model exists, but the tension is real. The flip side of a model that makes bold choices is that you have to steer it away from choices you did not ask for.

There is also the question every image company faces, which is where the training data came from and what happens to artists' styles. When TechCrunch asked Dorogush in 2024 whether an artist might sue over their style being used on the platform, she pointed to the terms of service for uploaded style references.

We are not using those to train our model.

Her argument was that the user supplies the style as a reference, so the act of borrowing a style belongs to the user, not the model. She also argued that tools like Recraft mostly replace bland stock imagery rather than designers, giving, say, a children's book author beautiful illustrations where they once would have settled for stock photos. Skeptics would note that a promise about user uploads is narrower than a full account of what the base model learned from in the first place, which is the question the whole industry keeps getting asked.

And competitors see the field differently. Nicolas Heymann, chief executive of the Berlin design platform Kittl, told TechCrunch in 2024 that if Recraft could reliably produce design assets, it might take market share from icon libraries, the same way image models were eating into stock photography. Dorogush answered that there was room for more than one player, and that Kittl focused on merchandise while Recraft focused on designers and marketers who need style consistency. Two years on, the bigger threat is not Kittl. It is the giants getting good at the very things Recraft built its lane around.

When to Reach for Recraft

Here is the practical part. When should you actually open Recraft, and when should you open something else? Let's go through it by job.

First, and most clearly, when you need a real vector file. If the output is a logo, an icon, a pictogram, or an illustration that someone will edit, scale, or send to a printer, Recraft is the default. Recraft claims it is the only model family that generates native editable SVG files from a prompt, and independent reviewers agree this is its strongest structural advantage. Version 4.1 Vector costs about eight cents per image through the interface for developers, and the Pro Vector version about thirty cents. If you already have a raster image you love, from any model, Recraft's vectorizer converts it for about one cent. And now you can fix the result node by node inside Studio.

Second, when a set has to match. An icon family, a series of illustrations for a recurring newspaper section, a campaign with twelve social posts. That is what version 4 Styles is for. Define the style once from references and keep generating. This is where the general purpose models still tend to drift, and where Recraft has put most of its engineering attention.

Third, when typography is part of the composition rather than a caption slapped on top. Posters, covers, album art, event flyers. Version 4 and 4.1 treat lettering as a structural element. The main rival here is Ideogram, which built its own reputation on text rendering, and it is worth testing both on your actual brief. Conveniently, Studio's compare switch lets you do exactly that in one click.

Fourth, when the output is going to print. The Pro versions output at higher resolution, and Recraft offers custom DPI settings, dots per inch, plus CMYK export, the cyan, magenta, yellow and black ink model that printing presses use. Most AI image tools live entirely in the screen world of RGB, red, green and blue light, and leave print conversion to you.

Fifth, for clean product shots and mockups, use version 4.1 Utility, not the main model. You want front facing, evenly lit, unsurprising. That is Utility's whole job, at about three and a half cents per standard image, or twenty-one cents for the Pro version at two thousand pixels. The mockup tools in Studio will then place a design onto a t-shirt, a mug, a sign or a package.

Sixth, when you only have a few words and want something that looks art directed. Wristwatch, metal and glass. Couple hugging. That is where the main version 4.1 model shines. And seventh, when you are sketching ideas and speed matters more than polish, try version 4.1 Flash, then refine only the drafts worth keeping.

Now the other side. When should you reach for something else? If the job is editing an existing photo conversationally, changing the jacket, moving the person, fixing the sky, or keeping a real person or a recurring character consistent across many photos, the strongest tools right now are Google's Nano Banana models and OpenAI's GPT Image models. They lead the general leaderboards for a reason, and their deep language understanding helps with complex multi-step instructions. The same goes for dense explanatory graphics, diagrams packed with labels and facts that have to be right. Models backed by large language models tend to handle that better.

If you want full local control, open weights, your own pipeline on your own machine, Recraft is not that. It is a closed, hosted service with an application programming interface. Open models like Flux or Qwen Image running in a node based tool such as ComfyUI give you that kind of tinkering freedom, at the price of doing the plumbing yourself. And if you need video, Recraft Studio can host video models from other companies, but Recraft itself does not make one.

One last workflow tip, which might be the most valuable of all. Because Studio hosts so many engines, you do not have to choose a single model religiously. You can generate a concept in whichever model gets the idea right, then pull it into Recraft's vector, style and print tools to make it production ready. And if you work with an AI assistant, Recraft's MCP server lets that assistant call Recraft directly, so the logo or the icon can come out of the same conversation where you wrote the copy.

Bring It Home

Let's make that concrete with the kind of small business work that never shows up in launch posts.

Picture a local newspaper that sells ads to the shops in its town. The café owner books a quarter page and emails over her logo. It is a blurry JPEG from 2009, 200 pixels wide, with a faint white box around it. The old answer was to apologize and redraw it, or print it fuzzy. The Recraft answer is to vectorize it for a cent, or regenerate a clean wordmark in version 4.1 Vector, nudge two curves in the editor, and drop a file into the layout that scales from a tiny classified to a full width banner without losing an edge. Export it in CMYK, and the printer is happy too.

Now picture the same paper running a recurring section that needs a small illustration every week. Instead of a different look every issue, you define a style once with version 4 Styles, and every new drawing looks like it came from the same hand. Or picture a rental company that needs clean, front facing shots of its equipment for a flyer. That is a Utility job, quick and predictable, no moody lighting required.

And that is the larger story underneath Recraft. The giant labs are racing to build the model that pleases the most people at everything, and they are very good at it. Recraft made a narrower bet. Not the picture that wins a blind vote, but the asset that survives a print shop, a brand guide, and the eighth image in the series. It is not so different from CatBoost. Not the flashy demo, but the dependable engine that working people quietly build on.

Whether that lane stays open is the real question for the next year. If OpenAI or Google ship truly good native vector output, a lot of Recraft's moat narrows to its workspace and its taste. But for now, if the thing you are making has to be a file that a designer can open, edit, and send to a press, the red panda is still the animal to call.

And to the other narrator who got this exact same brief today, may the best panda win.