A creative benchmark in eight prompts

Six models.
Eight runs.
Eight prompts.

A poster, an interface, a critique, a title sequence, an explainer, a story, a parser and a code review. Here is what came back.

01

What the benchmark found

Three things matter more than a ranking.

Technical

Code review separated the field.

Most runs found the two planted defects in the limiter. DeepSeek’s thick review missed the bigger one: prune compares an object to a number, so it never runs. It caught the reset bug, but not the type mismatch.

See the two reviews
Fidelity

Two runs are not directly comparable.

Muse Spark’s saved files were rewritten, not copied: four artifacts in run 1 and seven in run 2. Treat those runs as process demos, not as the model’s one-shot output.

See a trimmed artifact beside a clean one
Process

The cleanest runs were the quietest.

HY4 Flash and Claude Haiku 5.5 (run 2) used no subagent tools and saved every file exactly. DeepSeek’s subagents made 363 tool calls, browser screenshots included. GPT’s loaded 29 skills.

Browse the clean run
02

The runs

Eight runs across six model families. The small line under each name is the run caveat.

Run 1 Tool-assisted

DeepSeek 4.1 Flash

Subagents used a browser to refine their work.

363 tool calls 19m 31s slowest 2m 39s median 118 tok/s median 16/16 saved
View results
Run 1 One repair

GLM 5.3

One malformed CSS line was repaired after saving.

3 tool calls 12m 11s slowest 4m 24s median 102 tok/s median 16/16
View results
Run 1 Skill-assisted

GPT 6 Luna

Subagents loaded skills; two files were edited after saving.

30 tool calls 1m 33s slowest 37s median 73 tok/s median 14/16 exact
View results
Run 1 One unfenced reply

Claude Haiku 5.5

One review arrived without a fenced block.

0 tool calls 42s slowest 16s median 210 tok/s median 16/16
View results
Run 2 Clean run

Claude Haiku 5.5

No subagent tools; one identical retry.

0 tool calls 32s slowest 13s median 214 tok/s median 16/16 exact
View results
Run 1 Clean run

HY4 Flash

No subagent tools; all 16 files byte-exact.

0 tool calls 1m 29s slowest 31s median 175 tok/s median 16/16 exact
View results
Run 1 Not comparable

Muse Spark 1.3

Four visual files were saved trimmed.

5 tool calls 1m 06s slowest 38s median 88 tok/s median 12/16 exact
View results
Run 2 Not comparable

Muse Spark 1.3

Seven files saved trimmed; three prompts re-run after wording errors.

1 tool calls 1m 15s slowest 45s median 96 tok/s median 9/16 near-exact
View results
04

The eight tasks

Every task has a thin brief and a thick brief. Open a task to see all sixteen results.

How to read this

Creative work is presented. Technical work gets a note.

There is no invented score for a poster, an interface or a story. The parser and code-review tasks carry a short editorial note about what worked, what broke, or what missed the brief.

Read the method and caveats