Vyb: Coco Labs' First Design Model Builds Marketing Graphics in Editable Layers

Coco Labs announced Vyb on 20 October 2026: a design model that builds marketing graphics element by element instead of painting one flat image. The pitch is narrow and concrete. Give Vyb a brand's website or asset files once, and every design it produces comes back as a stack of editable layers — the headline is a text object you can retype, the logo is a vector, the product photo is the original pixels at full resolution. Change the headline and nothing else in the file moves. Design tooling launches rarely publish how the pipeline fails in production, and the recorded incidents are catalogued in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

That last sentence is the whole product. Anyone who has tried to fix a typo in an AI-generated ad knows the alternative: re-run the prompt, get a slightly different design, and lose the layout you had already approved. Vyb is Coco Labs' attempt to stop that cycle, and it is the company's first model release of any kind.

What actually shipped

The launch post is by David Zats and Anusha Ramesh and describes Vyb as "the first design model that understands a user's taste." The mechanism matters more than the slogan. Vyb composes a design in an internal format Coco Labs calls Cocolang — a terse, structured description of the design that both humans and models can read and edit. Text, vector shapes and images keep their native formats inside that structure rather than being flattened into one raster. Because the output is structured, it can be exported "losslessly" into external design tools such as Figma, where each element arrives on its own layer.

Vyb can also call image generators and editing tools when a job needs something it cannot draw itself, which is the sensible division of labour: generate the photograph, place it as a layer, keep the type and layout crisp around it.

What did not ship is nearly as important. There is no rate card. There is no public API. There is no published parameter count, no architecture description and no training-data disclosure beyond the fact that Coco Labs built a proprietary pipeline for layered designs because few public layered designs exist to train on. Access is a waitlist, and joining it comes with $10 in generation credits. As of writing, the product is announced and invite-only, not generally available.

The three numbers Coco Labs is leading with

Coco Labs published a benchmark it calls VybBench, covering 145 G​oogle fonts, 100 colours and 200 edits against three rivals: ChatGPT Image (gpt-image-2.5), Muse (muse-image) and Recraft v4.1. All of the figures below are vendor-reported. Nobody outside Coco Labs has run them, and the benchmark's author is the company whose model wins it.

• Typeface fidelity — shape-match score against a programmatic rendering of the same word: Vyb 90.3%, ChatGPT Image 76.0%, Muse 50.9%, Recraft 46.2%. Coco Labs measured a noise ceiling of 94.4% by scoring identical renderings against each other, so 90.3% is close to the practical maximum on this test.

• Colour fidelity — mean maximum per-channel error out of 255, lower is better: Vyb 0.00, ChatGPT Image 3.94, Muse 5.67, Recraft 27.19. A literal zero is possible here because a colour can be written as a value rather than painted; that is the layered format doing the work, not a better painter.

• Edit locality — mean share of pixels changed across 200 edit instructions: Vyb 11.9%, ChatGPT Image 95.5%, Muse 91.8%, Recraft 98.9%. An independent LLM judge also scored whether each requested edit was actually completed: Vyb 94.7%, ChatGPT Image 100%, Muse 98.0%, Recraft 50.3%.

Read the third row carefully, because it is the one that decides whether the product is real. A pixel-diff model cannot change 11.9% of an image by accident; it changes almost everything because it regenerates almost everything. Vyb changes little because it is editing a document. The 94.7% acceptance rate next to a 100% from ChatGPT Image is the honest cost of that: Vyb occasionally refuses an instruction it judges to be bad design, such as a change that would make text unreadable, and the judge counted the refusal as a miss.

Coco Labs discloses one methodological asymmetry in its own benchmark: Vyb received layered SVG inputs for the editing test while the other three received flattened PNGs, because they cannot accept layers. That is not a trick — it is the difference being measured — but it does mean the comparison tests two different input contracts, not two models given identical work.

The argument underneath the benchmark

Vyb's launch post spends a paragraph attacking a rival approach, and it is worth quoting the shape of it rather than the wording. The claim is that converting a finished flat image back into layers — de-layering — cannot recover genuine text or vector layers, because the text in a flat image is only pixels, and any region hidden behind another element has to be guessed at. Coco Labs is arguing that the only way to get a real layer stack is to generate one, and that post-hoc decomposition is a demo that does not survive contact with a brand's actual files.

That is a defensible position, and it is also a marketing position from the company selling the alternative. It is true that text rasterised into a flat image cannot be re-typed with perfect fidelity to the original font metrics. It is also true that most marketing teams already have a library of approved flat assets they would love to make editable, and Vyb has nothing to offer them for those files. The two approaches solve different problems, and the launch post is written as though only one of them counts.

Where it lands in a real workflow

A layered design model does not remove the rest of the pipeline. Someone still has to write the headline, decide the offer, and check that the claim is legal in the market it runs in. What Vyb changes is the step after that: the copy is placed as text, in the brand's font, in a layout that can be adjusted by a designer rather than re-rolled by a prompt.

The text half of that pipeline is also the half a router can serve. OrcaRouter runs one API across 200+ models with provider list prices passed through at 0% markup, so the model writing the headline and the model checking it against brand rules can sit behind a single key without a second contract. Automatic failover matters here for a boring reason: a copy-generation step that fails mid-batch leaves a design job with a blank layer, and a fallback model costs less than a stalled queue. Vyb itself is not callable through any router — it has no public API — so the honest architecture in late 2026 is a text model for the words and Vyb, once you have access, for the artwork.

What is still unproven

Four things, in descending order of how much they should affect a buying decision.

• Price. No rate card exists. Coco Labs is handing out $10 in credits to waitlist joiners, which tells you nothing about what a finished campaign costs.

• Independent testing. Every number in this article comes from the vendor. The font benchmark covers publicly available G​oogle fonts only; Coco Labs says it expects the gap to widen for custom brand fonts but did not measure those, which is precisely the case a brand cares about.

• File fidelity in practice. "Lossless export to Figma" is a claim about the exporter, not about whether the resulting layer tree is one a designer wants to work in. Nobody outside the company has opened one of these files.

• Coverage. The launch shows posters, ads, quote cards and announcements. Multi-page documents, packaging dielines and anything with a print colour profile are unaddressed in the material published so far.

None of that makes the product vapour. A design model whose output is a document rather than a picture is a genuinely different category from an image generator, and the edit-locality number is the kind of result that only comes from that architecture. But a waitlist and a vendor benchmark is an announcement, not a track record, and the useful question for a marketing team is not whether Vyb beats ChatGPT Image at fonts — it is whether the files it produces save a designer more time than re-doing the layout by hand would have cost. That question gets answered when the waitlist opens, and not before.

What to watch

Three signals would move Vyb from interesting to adopted. A published price per design or per credit, so the cost can be compared against a freelancer's hourly rate. A custom-font benchmark, because every real brand uses at least one typeface that is not on G​oogle Fonts. And an API, because the moment Vyb is callable it stops being a tool and becomes a component that can be routed alongside everything else in a content pipeline. Until then, the $10 credit is the cheapest way to find out whether the layer tree is as clean as the benchmark implies.