/
Design System Bench

Design System Benchv1

Design System Bench measures whether coding agents build new UI from a repo's own design system or write their own markup instead.

Nov 16, 2026 · by 21st · 4 min read

We gave 14 models a short ticket from a real product's repo and asked each one to build the UI it describes. 20 tasks from 6 open-source products with their own design systems: Cal.diy, Supabase, PostHog, Trigger.dev, Polar and Plane.

Fable 5.2 in Claude Code scored highest, a few points ahead of Gemini 4 Carbon and GPT-6 Astra. Every model still hand-builds some of what the repo already has, mostly buttons and badges, and the middle of the table is close to a tie.

How it works

The tasks

10 tasks where we deleted one component file and the agent writes it again, and 10 where it rebuilds a merged pull request. The human's version is hidden.

What models get

The whole repo and a 2 to 4 sentence ticket, the way people really ask. It doesn't say which components to use.

Ticket · Trigger.dev DeliveriesTable
Add `DeliveriesTable` in `apps/webapp/app/components/webhookDeliveries/v1/DeliveriesTable.tsx` to list a webhook endpoint's deliveries on the endpoint detail page, so users can scan status and errors at a glance. Every row should link through to that delivery's detail page, and failed deliveries should show their error message, with the full text available on hover.
The human's change
<TableRow key={delivery.id}>  <TableCell to={deliveryPath}>    <span className="flex items-center gap-1.5">      <TruncatedCopyableValue value={delivery.friendlyId} className="text-xs" />      {delivery.isTest ? <Badge variant="extra-small">Test</Badge> : null}    </span>  </TableCell>  <TableCell to={deliveryPath}>    <DeliveryStatusBadge status={delivery.status} />  </TableCell>  <TableCell to={deliveryPath}>    {delivery.externalDeliveryId ? (      <div className="w-[24ch]">        <MiddleTruncate          text={delivery.externalDeliveryId}          className="font-mono text-xs"        />      </div>    ) : (      <span className="text-text-dimmed group-hover/table-row:text-text-bright">        None      </span>
The real ticket, and the human's change the agent never sees.

The models

Each runs in its maker's own agent: Claude Code, Codex, Gemini CLI, Grok Build and so on, at medium thinking.[1]

One run, 60 min

One run per task in a sandbox. Running out of time counts, and is graded on whatever the agent wrote.

Scoring

Of the design-system components the human used, the share the agent used too, counted over the parts it built at all. Another design-system component a maintainer would accept also counts.[2]

Checking the grader

The human's own change scores 100 on all 20 tasks, and every trace shows the intended model answering.

Results

Leaderboard

1Fable 5.2Claude Code0.0
2Gemini 4 CarbonGemini CLI · high thinking0.0
3GPT-6 AstraCodex0.0
4Qwen 3.8 Max PrimeQwen Code0.0
5GPT-6.1 SolCodex0.0
6MiniMax M3MiniMax Codeopen weights0.0
7GPT-6 SolCodex0.0
8Opus 5.5Claude Code0.0
9Grok 4.7Grok Build0.0
10GLM 5.3Claude Code0.0
11Muse WonderMuse Code0.0
12Sonnet 5.5Claude Code0.0
13Kimi K3.1 ProKimi Codeopen weights0.0
14MiMo V2.6 ProMiMoopen weights0.0
Time per task
Minutes per task, fastest first
42s
1.2
1.3
1.4
1.5
1.6
1.7
1.9
2.5
6.4
10.0
11.5
28.8
32.0
Cost per task
Dollars per task, cheapest first. * Estimated
$0.14*
$0.19
$0.20
$0.31
$0.31*
$0.64
$0.65
$0.68*
$0.89*
$0.99
$1.03*
$1.21
$1.31*
$2.34*

The faint band is the 95% interval over tasks. Cost and time are per task. * Tokens times the host's price: the agent doesn't report cost.

What they built

Each change rendered in the product's own app: the human's first, then every model's, best score first.[5]

Ticket Add EditForm (named export) in apps/web/modules/users/components/UserTable/EditSheet/EditUserForm.tsx so org admins can edit a member from the users table side sheet: avatar, name, username, bio, role and timezone. The Owner role option should only appear when the signed-in user is an org owner. For now, submitting just exits edit mode; the server call comes later.

The human's changehidden from the agents
The human's change, rendered
Every model, best score first
The human's change, rendered
The human
the reference

2 didn't render. 10 hand-built at least one part the repo has a component for. Score is reuse on built parts for this task. Click a tile to compare it with the human's.

Insights

Most of the time the agent finds the component

Over 280 runs the human's changes used 3,458 design-system components. Agents used the same one 61% of the time and another design-system component 15%. They hand-built 7%.

What the agent did with the human's component
Used the same component0%
Used another design-system component0%
Didn't build that part0%
Hand-built it with raw markup0%
Wrote a new local component0%
Not needed for its version0%
Every design-system component in the human's 20 changes, over all 14 models. A second model sorts the misses.

What gets hand-built: buttons and badges

Button (27 runs) and Badge (19) top the list. On one Trigger.dev table, 13 of 14 models wrote a styled span for a "Test" tag the repo already has as <Badge>.

Plane is the hardest repo

The average model scores 87 on Supabase and 63 on Plane. Plane's worst task is the shortcuts dialog, where 11 of 14 models wrote their own search input.

Reuse on built parts0100
ModelSupabase3 tasksTrigger.dev3 tasksCal.diy3 tasksPolar4 tasksPostHog3 tasksPlane4 tasks
Fable 5.294.087.782.391.885.772.0
Gemini 4 Carbon88.091.080.383.090.369.2
GPT-6 Astra91.086.780.091.266.765.0
Qwen 3.8 Max Prime89.782.077.389.051.380.0
GPT-6.1 Sol89.785.780.079.285.757.8
MiniMax M386.383.386.780.566.364.0
GPT-6 Sol87.783.380.080.071.054.2
Opus 5.583.381.777.378.274.752.2
Grok 4.782.781.069.057.579.371.8
GLM 5.385.080.749.374.571.068.8
Muse Wonder82.382.777.380.242.055.0
Sonnet 5.576.780.377.369.564.749.0
Kimi K3.1 Pro91.380.369.052.558.364.5
MiMo V2.6 Pro90.785.775.048.847.361.7
Average of 1487.083.775.875.468.263.2
Reuse on built parts per model and repo. Repos have 3 to 4 tasks, so a single cell is mostly noise; read the columns.

More time doesn't buy more reuse

Fable 5.2 scored 85 in 1.4 minutes a task; Gemini 4 Carbon took 32 for 83, and MiMo V2.6 Pro 29 for 67.

Nobody builds the whole ticket

Models build 44 to 59% of the hidden checklist, the human 98%. The ticket leaves most of it unsaid, so that's a measure of scope, not quality.[3]

The agent around the model matters

We first ran every model in the same agent, mini-SWE-agent. GPT-6 Astra scored 69 there and 80 in Codex; GPT-6.1 Sol 65 and 79. The two runs used different thinking levels, so this isn't a clean comparison.[4]

1Gemini 4 Carbonmini-SWE-agent0.0
2Opus 5.5mini-SWE-agent0.0
3Fable 5.2mini-SWE-agent · low thinking0.0
4Grok 4.7mini-SWE-agent0.0
5Kimi K3mini-SWE-agent · K3, not K3.1 Pro0.0
6Muse Wondermini-SWE-agent0.0
7GPT-6 Astramini-SWE-agent0.0
8Sonnet 5.5mini-SWE-agent0.0
9GLM 5.3mini-SWE-agent0.0
10GPT-6.1 Solmini-SWE-agent0.0
11GPT-6 Solmini-SWE-agent0.0
The same 20 tasks, every model in mini-SWE-agent at its provider's default thinking (Fable 5.2 at low). A separate run: don't subtract it from the leaderboard.

Notes

  1. [1]

    Gemini 4 Carbon ran at high thinking (Gemini CLI has no medium for it); every other model ran at medium. DeepSeek V4 Pro isn't in the table; our API account blocks providers that train on prompts, and that blocked it. Costs marked * are tokens times the host's price, because those agents don't report cost.

  2. [2]

    A second model decides whether a different component would be accepted. One run per task on 20 tasks only separates gaps of about 10 points: the top (Fable 5.2 85, Gemini 4 Carbon 83, GPT-6 Astra 80) and the bottom (Kimi K3.1 Pro 68, MiMo V2.6 Pro 67) differ, and the middle is close to a tie.

    The score doesn't check that the code builds: in the renders, GPT-6 Astra and GPT-6.1 Sol don't build on Cal.diy's edit-user form (they import a path the package doesn't export) yet scored 85 and 85 on it, and Kimi K3.1 Pro doesn't build on Supabase's upgrade alert.

  3. [3]

    A blind judge marks each checklist item against the diff. The checklist is written from the full spec, so it includes exact labels the ticket never mentions.

  4. [4]

    The first experiment ran each model at its provider's default thinking (GPT at medium, Claude at high, Kimi and GLM at max), used Kimi K3 rather than K3.1 Pro, and Fable 5.2 at low thinking. So comparing it with the second experiment, each model in its own agent, isn't a clean comparison.

  5. [5]

    Four of the six repos (Trigger.dev, Cal.diy, PostHog and Plane) were rendered on a Vite page inside the repo with its real styles and design-system packages, not the full app with its server.