VibeTestModel Evaluation Archive
中文
ARCHIVE 001 / 2026 · Exhibition season

VibeTest

An archive of things AI models actually made.

记录模型真正创造出来的东西。

On the same task set, compare what each model actually made, how the run went, and what independent reviewers concluded.

Browse tasks →Compare works →
  • Same prompts
  • Real outputs
  • Independent reviews
  • Verifiable evidence
02 · SVG
03 · SVG
AI CREATIVITY IN THE OPENSame prompts · Real outputs · Independent reviews · Verifiable evidence
04ModelsSame task set
02PhasesPhase 1 complete
70RunsEvery run kept
20 / 45TasksArchived / planned
Archive progress · Tested / planned20 / 4544.4%
Denominator
Sum of planned tasks
Task versions
20
Total runtime
11:08:01
FEATURED WORKS

Featured works

All three results use the same width and the same preview ratio; no model is marked as best.

View all tasks →
CURRENT EXHIBITION

Current exhibition

Phases →
01
PHASE-01

Phase 1 · 15 benchmark tasks

SVG illustration, single-file interactive apps, and multi-file web applications.

15 / 15 Tasks · Completed
4 Models · 60 Runs
02
PHASE-02

Phase 2 · advanced creation tasks (Tasks 16–45)

In-browser animation, simulation, engineering and authoring tools. This archive currently covers only the tested and archived Tasks 16–20 (5 tasks × 2 models = 10 runs); the remaining tasks have not been run.

5 / 30 Tasks · In progress
2 Models · 10 Runs
MODELS IN ARCHIVE

Models in archive

All models →
M01
DeepSeek

DeepSeek-V4.1-Flash (0910 preview + GA release)

20 works · 5:24:18

→M02
Meta Superintelligence Labs

Muse Spark 1.3

20 works · 1:10:47

→M03
OpenAI

GPT-5.6 Sol

15 works · 1:16:46

→M04
Moonshot AI（Kimi）

K3

15 works · 3:16:10

→
RECENTLY ARCHIVED

Recently archived

The archive does not record a per-run archive date, so these are listed by phase 02 and archive order (group runtime 3:15:56).

Phase details →
METHODOLOGY

Methodology

Read the methodology →
01

Same prompts

Models in one phase receive exactly the same prompt text and resource limits. Prompt versions increase instead of overwriting the original, and comparison stays within one task version.

02

Real outputs

Pages show the files the models actually delivered: pages can be opened, SVGs can be inspected, and run records keep the call and retry counts.

03

Independent reviews

Human reviews, each AI review version and this website’s own checks are recorded separately; scores are shown side by side with their method and disclaimer, never silently merged.

04

No aggregate ranking

Scores from different reviewers and rubrics are never merged into a leaderboard; unknown values stay "Not recorded" instead of being inferred.