Design and build Canvely's responsive design editor from the frozen prompt, rules, document data, and requirements, including zoom around the pointer, shape-accurate selection, snapping, rotated resize, groups, and undo history without a canvas, drag-and-drop, or geometry library.
The v2 rules inside one identical sandbox for every tool, with a provided headless Chromium for the agent’s own localhost preview. Runs are compared only within one protocol. How the protocols differ
One archived screenshot per published Run against version 1.0.0, of the Run’s representative Revision. Ordered by setup and then by repetition, not by quality.
Frozen with this version. Requirement IDs are stable across versions and are cited unchanged by evaluation records.
Automated checks are not an accessibility certification. Recorded axe findings describe what the tooling observed at capture time, on a pinned browser and viewport. A check that reports no violation is not proof that the interface is accessible, and a selector problem is inconclusive rather than an implementation failure.
| # | Model | Tool | Revision | Status | Actions |
|---|---|---|---|---|---|
| Fig. 01 | Claude Opus 5.5 (high) | Claude Code | 0 · initial | live | |
| Fig. 02 | GPT-6-Luna (max) | Codex CLI | 0 · initial | live | |
| Fig. 03 | GPT-6-Sol (high) | Codex CLI | 0 · initial | live |
Fig. 01Claude Opus 5.5 (high)
Fig. 02GPT-6-Luna (max)
Operator-attested: the Configuration, tool version, displayed model, and elapsed time were recorded by the operator from what the tool showed at Run time. They are not independently verified. An exact model ID appears only when the tool exposed one; it is never inferred.
Published sites are independent, AI-generated demos served from separate preview origins with local mock data. They are not products, and no UIBenchmark session data crosses to them.
Exact model ID, tool version, elapsed time, run state, stop reason, discarded attempts and every published Revision, repairs included, in run log order. Each Run and each Revision links to its own page.
01Run 1 · Claude Opus 5.5 (high) · Claude Code
These Runs were executed against version 1.0.0, the most recently frozen Challenge Version under protocol v3. Runs against another frozen Version are listed on that Version’s own page, because they received different inputs.
Design and build Canvely's responsive design editor from the frozen prompt, rules, document data, and requirements, including zoom around the pointer, shape-accurate selection, snapping, rotated resize, groups, and undo history without a canvas, drag-and-drop, or geometry library.
Fictional browser-based design tool, vector editor with canvas, layers, and properties
The workflow designs the interface itself from the frozen brief.
Read PROMPT.md and BENCHMARK_RULES.mdThis is the brief of the most recently frozen Challenge Version under protocol v3. The full text of its launch prompt and shared rules, and the frozen inputs of earlier versions, are on the Challenge Version pages, so nothing here is a paraphrase of what a Run actually received.
Newest first. A frozen version can never change; a correction becomes a new version.
The declared rules and viewports version 1.0.0 was executed and captured under.
Fig. 03GPT-6-Sol (high)
02Run 1 · GPT-6-Luna (max) · Codex CLI
03Run 1 · GPT-6-Sol (high) · Codex CLI
Discarded attempts failed for a technical reason — a crashed session, a tool outage, a broken upload — and were discarded and retried from the same frozen inputs. The count is not a quality signal.
Generation cost: Not separately metered. Runs were executed on the operator’s personal tool subscriptions, which do not itemize a per-Run charge. No amount is attributed, and no zero-cost claim is made.
Starter SHA-256
aa2276a44780653d1147ac1eb10925e101c37b45fe6406afc268fc2d15db4a73
Asset pack SHA-256
4794e27a536d59f148e0ab852c3621dd269495ace6fe92b47a826013140cb278