AI coding agent benchmark
What AI coding agents build. Same prompt, side by side.
60 runs from 12 models on one lava lamp prompt, checked by the same hidden browser tests. Every run is published.
Results so far · Lava lamp v1
Every model at every effort. 60 published runs, 12 models.
Test runs before the first episode. Click a build to open it.
| Model | Low12 models | Medium12 models | High12 models | Extra high12 models | Max12 models |
|---|---|---|---|---|---|
| Claude Code | |||||
| Claude Fable 5.15 runs · 5 of 5 work |  Works 2m 50s 1 run |  Works 4m 53s 1 run |  Works 14m 06s 1 run |  Works 32m 58s 1 run |  Works 36m 52s 1 run |
| Claude Haiku 5.55 runs · 5 of 5 work |  Works 2m 36s 1 run |  Works 5m 06s 1 run |  Works 5m 22s 1 run |  Works 8m 28s 1 run |  Works 20m 00s 1 run |
| Claude Opus 5.55 runs · 5 of 5 work |  Works 1m 42s 1 run |  Works 8m 56s 1 run |  Works 13m 48s 1 run |  Works 28m 09s 1 run |  Works 66m 39s 1 run |
| Claude Sonnet 5.55 runs · 5 of 5 work |  Works 2m 37s 1 run |  Works 4m 44s 1 run |  Works 6m 45s 1 run |  Works 21m 18s 1 run |  Works 87m 03s 1 run |
| Codex CLI | |||||
| GPT-5.6 Luna5 runs · 5 of 5 work |  Works 29.7 sFastest 1 run |  Works 59.1 s 1 run |  Works 1m 04s 1 run |  Works 1m 25s 1 run |  Works 2m 22sFastest 1 run |
| GPT-5.6 Sol5 runs · 2 of 5 work |  Works 1m 46s 1 run |  Fails 2m 20s 1 run |  Works 2m 34s 1 run |  Fails 4m 05s 1 run |  Fails 7m 53s 1 run |
| GPT-5.6 Terra5 runs · 4 of 5 work |  Works 1m 38s 1 run |  Works 1m 48s 1 run |  Works 1m 00sFastest 1 run |  Works 3m 50s 1 run |  Fails 4m 04s 1 run |
| GPT-6 Astra5 runs · 4 of 4 work | No capture published Not checked 4m 17s 1 run |  Works 4m 15s 1 run |  Works 7m 30s 1 run |  Works 25m 55s 1 run |  Works 36m 32s 1 run |
| GPT-6 Luna5 runs · 4 of 5 work |  Fails 51.5 s 1 run |  Works 1m 21s 1 run |  Works 1m 27s 1 run |  Works 4m 24s 1 run |  Works 4m 38s 1 run |
| GPT-6 Sol5 runs · 5 of 5 work |  Works 2m 39s 1 run |  Works not recorded 1 run |  Works 4m 12s 1 run |  Works 7m 11s 1 run |  Works 13m 48s 1 run |
| GPT-6.1 Sol5 runs · 5 of 5 work |  Works 3m 11s 1 run |  Works 4m 52s 1 run |  Works 5m 15s 1 run |  Works 11m 37s 1 run |  Works 16m 56s 1 run |
| GPT-6.1 Terra5 runs · 5 of 5 work |  Works 42.9 s 1 run |  Works 47.9 sFastest 1 run |  Works 1m 12s 1 run |  Works 1m 15sFastest 1 run |  Works 4m 55s 1 run |
Low effort · 12 models · fastest first
GPT-5.6 Luna29.7 s
Works
GPT-6.1 Terra42.9 s
Works
GPT-6 Luna51.5 s
Fails
GPT-5.6 Terra1m 38s
Works
Claude Opus 5.51m 42s
Works
GPT-5.6 Sol1m 46s
Works
Claude Haiku 5.52m 36s
Works
Claude Sonnet 5.52m 37s
Works
GPT-6 Sol2m 39s
Works
Claude Fable 5.12m 50s
Works
GPT-6.1 Sol3m 11s
Works
- No capture published
GPT-6 Astra4m 17s
Not checked
Medium effort · 12 models · fastest first
GPT-6.1 Terra47.9 s
Works
GPT-5.6 Luna59.1 s
Works
GPT-6 Luna1m 21s
Works
GPT-5.6 Terra1m 48s
Works
GPT-5.6 Sol2m 20s
Fails
GPT-6 Astra4m 15s
Works
Claude Sonnet 5.54m 44s
Works
GPT-6.1 Sol4m 52s
Works
Claude Fable 5.14m 53s
Works
Claude Haiku 5.55m 06s
Works
Claude Opus 5.58m 56s
Works
GPT-6 Solno time
Works
High effort · 12 models · fastest first
GPT-5.6 Terra1m 00s
Works
GPT-5.6 Luna1m 04s
Works
GPT-6.1 Terra1m 12s
Works
GPT-6 Luna1m 27s
Works
GPT-5.6 Sol2m 34s
Works
GPT-6 Sol4m 12s
Works
GPT-6.1 Sol5m 15s
Works
Claude Haiku 5.55m 22s
Works
Claude Sonnet 5.56m 45s
Works
GPT-6 Astra7m 30s
Works
Claude Opus 5.513m 48s
Works
Claude Fable 5.114m 06s
Works
Extra high effort · 12 models · fastest first
GPT-6.1 Terra1m 15s
Works
GPT-5.6 Luna1m 25s
Works
GPT-5.6 Terra3m 50s
Works
GPT-5.6 Sol4m 05s
Fails
GPT-6 Luna4m 24s
Works
GPT-6 Sol7m 11s
Works
Claude Haiku 5.58m 28s
Works
GPT-6.1 Sol11m 37s
Works
Claude Sonnet 5.521m 18s
Works
GPT-6 Astra25m 55s
Works
Claude Opus 5.528m 09s
Works
Claude Fable 5.132m 58s
Works
Max effort · 12 models · fastest first
GPT-5.6 Luna2m 22s
Works
GPT-5.6 Terra4m 04s
Fails
GPT-6 Luna4m 38s
Works
GPT-6.1 Terra4m 55s
Works
GPT-5.6 Sol7m 53s
Fails
GPT-6 Sol13m 48s
Works
GPT-6.1 Sol16m 56s
Works
Claude Haiku 5.520m 00s
Works
GPT-6 Astra36m 32s
Works
Claude Fable 5.136m 52s
Works
Claude Opus 5.566m 39s
Works
Claude Sonnet 5.587m 03s
Works
Build time: the number is the median, each dot one run, on one scale per effort.
First episode: date not announced yet. How it will run
- 01
Same task
Same prompt, files and hidden checks for every agent.
- 02
Same conditions
Same deadline, same number of runs, shuffled order.
- 03
Checked in a browser
Every build, on a desktop and a phone screen.
- 04
Everything published
Every run, failed ones too, with captures and source.