Vyb vs Qwen 3.8: A Design Model Against the Reasoner That Would Brief It

Six tenths of a cent. That is what a 3,000-token design brief costs to reason over on Qwen3.8-Flash at its published $0.15 per million input tokens and $0.47 per million output — and it is the number that makes this comparison worth writing. Vyb, Coco Labs' design model announced on 20 October 2026, produces layered, editable marketing graphics from a brand's own assets but publishes no price and has no API. Q​wen 3.8, A​libaba's generation that went generally available on 3 August 2026, produces no pictures at all but reasons about text and images for a rate card you can read. One of these you can budget to the cent today. The other you cannot budget at all. Neither a design product nor a reasoning model records how the workflow fails once it ships, 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.

So this is not a head-to-head in the usual sense. It is a question about where the money and the planning go in a design pipeline, and about whether a model that thinks can substitute for a model that draws.

The Q​wen 3.8 side, precisely

Q​wen 3.8 is a generation, not a single model, and the member you pick changes the arithmetic by more than an order of magnitude.

• Qwen3.8-Flash — $0.15 per million input tokens and $0.47 per million output, a 1M-token context, and text, image and video input. This is the volume lane, and for brief-parsing it is almost free.

• Qwen3.8-27B — a dense 27-billion-parameter multimodal model under Apache 2.0, 262,144-token context extendable toward a million, 55.6 GB of BF16 weights across eighteen shards. Self-hostable, which means the marginal cost of a call is your electricity.

• Qwen3.8-Max — the 2.4-trillion-parameter sparse mixture-of-experts flagship, roughly 95B parameters active per token, $2.00 per million in and $6.00 per million out, with a 1M-token context.

All three read images. None of them draws one. Qwen3.8-27B and Flash take text, image and video as input and return text; the Max tier is the same contract at a higher price. That is the whole limitation this comparison turns on.

The arithmetic of the brief

A design job starts with a brief: the offer, the audience, the placements, the brand rules, the examples of what worked last quarter. In a pipeline, a reasoning model reads that and emits something structured — a layout plan, a copy deck, a set of constraints. That is work Q​wen 3.8 does well, and its cost is small enough to compute rather than estimate.

• Small brief — 2,000 input tokens, 600 output. On Qwen3.8-Flash: $0.00030 in plus $0.00028 out, or $0.00058. Under a tenth of a cent per design.

• Medium brief — 20,000 input, 600 output, which is enough for a full brand guide and a dozen approved examples. $0.0030 plus $0.00028, or $0.0033. A third of a cent.

• Large brief — 200,000 input, 600 output, the whole campaign archive in context. $0.030 plus $0.00028, or $0.030. Three cents, still.

• The same three jobs on Qwen3.8-Max — $0.0040, $0.0403 and $0.4036 respectively. Thirteen times the price at every size, and still under half a dollar for the largest brief you would realistically write.

Put those against Vyb's cost, which is $10 of free credits and then an unpublished number, and the asymmetry is stark. The reasoning half of a design pipeline is effectively free at any scale a marketing team operates at. The producing half is the entire cost, and it is the half nobody has priced.

Why the layer format is the real argument

Coco Labs spends a section of its launch post attacking de-layering — the practice of taking a flat image and decomposing it into layers after the fact. The argument is that a flat image's text is only pixels, so reconstructed text layers cannot carry the original font's true metrics, and anything hidden behind another element has to be guessed. Coco Labs' position is that a genuine layer stack has to be generated as a layer stack.

That argument is aimed at a real target. Q​wen's own family includes layered image work, and the industry's general direction has been to make existing images editable rather than to make editable images from scratch. Coco Labs is betting on the second approach, and its benchmark is built to prove it: typeface fidelity of 90.3% against 76.0% for ChatGPT Image, colour error of literally 0.00 against 3.94, and mean pixel change across 200 edits of 11.9% against 95.5%. Those are vendor-reported figures from a benchmark Coco Labs wrote, and the editing test gave Vyb layered SVGs while its rivals got flattened PNGs — an asymmetry the company discloses, because the layer format is precisely the thing being measured.

What the argument does not address is the corpus problem. Every brand already owns thousands of flat assets — last year's campaigns, the deck that legal approved, the hero image from the last launch. A generative-layer model cannot touch any of them. A de-layering approach at least attempts to. Vyb is better at making new editable designs; it offers nothing for the archive, and the launch post presents that as a reason the other approach is wrong rather than as a boundary.

Where Q​wen 3.8 actually earns its place in this pipeline

Three jobs, in ascending order of how much they matter.

Reading the brand. Vyb ingests a website and asset files; something has to decide what in that material is a rule and what is an accident. A vision-language model can read a page of past campaigns and extract the pattern — the typeface hierarchy, the colour that appears on every CTA, the recurring phrase — and hand that to a human for approval. Qwen3.8-27B does this at 262K context and under Apache 2.0, so a team can run it privately on brand material that should not leave the building.

Writing to the layer. Vyb's output is a document with a text object in it. Something has to fill that object with a headline that fits, scans and does not make a claim the legal team will reject. That is the Q​wen 3.8 job this page's cost section priced at six tenths of a cent.

Checking the result. A rendered design can be sent back to a multimodal model for a read: does the headline fit the safe area, is the logo legible against the background, does the contrast pass. Qwen3.8-Flash takes images as input and costs $0.15 per million tokens, which makes a per-asset visual check cheaper than the storage of the asset itself.

What a routed pipeline looks like

The practical shape in late 2026 is a text and vision layer that is fully automatable and a design layer that is not. Q​wen 3.8 sits entirely on the first side: brief in, structured plan out, check on the rendered result, all through one endpoint. Vyb sits on the second side behind a waitlist, with no API, which means a person drives it.

That split is exactly what a router is for on the automated half. OrcaRouter carries Qwen3.8-Flash at $0.15 and $0.47, Qwen3.8-27B at $0.33 and $2.40, and Qwen3.8-Max at $2.00 and $6.00 — the providers' own list prices passed through at 0% markup, so a vendor price cut lands on the same key the same day. A brief-parsing call and a rendered-design check can run against different tiers on one credential, with automatic failover between them, which is the difference between a batch that finishes and a batch that stops on the thirty-first asset. Vyb is not routable and this article will not pretend otherwise; there is no public endpoint to route to.

The budget question nobody can answer yet

Here is the thing a marketing director will ask, and here is the honest state of the answer.

"What does it cost to produce a campaign?" The reasoning side: between a third of a cent and three cents per brief on Flash depending on how much context you allow, plus a few tenths of a cent for the visual check. Call it a rounding error. The design side: unknown. Coco Labs has not published a per-design price, a subscription tier or a credit rate beyond the $10 waitlist grant. Until it does, the total cost of an automated design pipeline has one measured component and one blank.

That blank is not a reason to dismiss the product — the benchmark's edit-locality number is the kind of result that only comes out of a genuinely structured output format, and no reasoning model can produce it. But it is a reason to be precise about what Vyb is: a promising design component whose economics are unhidden, sitting next to a reasoning layer whose economics are public to four decimal places.

What would settle it

Two disclosures would turn this from an interesting contrast into a build decision. A Vyb price per design, so the pipeline total stops having a blank. And a custom-font benchmark, because Coco Labs' 145 G​oogle fonts are not the typefaces a real brand uses, and the company says itself it expects the gap to widen for custom faces without having measured them.

Until both land, the sequence that makes sense is the one the pricing already implies. Automate the reasoning, the brief-parsing and the visual check now, on a tier you can cost to the cent. Treat the design step as a pilot with $10 of credits and a designer watching, and decide on the full pipeline when the rate card appears. The six-tenths-of-a-cent number is real today; the rest is a waitlist.