AI coding agent benchmark
What AI coding agents build. Same prompt, side by side.
86 runs from 19 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. 86 published runs, 19 models.
Click a build to open its run.
| Model | Low18 models | Medium18 models | High18 models | Extra high19 models | Max13 models |
|---|---|---|---|---|---|
| Claude Code | |||||
| Claude Fable 5.15 of 5 work | |||||
| Claude Haiku 5.55 of 5 work | |||||
| Claude Opus 5.55 of 5 work | |||||
| Claude Sonnet 5.55 of 5 work | |||||
| Claude Fable 5.24 of 4 work | Not run | ||||
| Codex CLI | |||||
| GPT-5.6 Luna5 of 5 work | |||||
| GPT-5.6 Sol2 of 5 work | |||||
| GPT-5.6 Terra4 of 5 work | |||||
| GPT-6 Astra5 of 5 work | |||||
| GPT-6 Luna4 of 5 work | |||||
| GPT-6 Sol5 of 5 work | |||||
| GPT-6.1 Sol5 of 5 work | |||||
| GPT-6.1 Terra5 of 5 work | |||||
| GPT-6.2 Astra4 of 4 work | Not run | ||||
| GPT-6.2 Sol4 of 4 work | Not run | ||||
| OpenCode | |||||
| Muse Ember5 of 5 work | |||||
| Inkling 2 Mini4 of 4 work | Not run | ||||
| Qwen3.8 27B1 of 1 work | Not run | Not run | Not run | Not run | |
| Antigravity CLI | |||||
| Gemini 4 Neon4 of 4 work | Not run | ||||
Low effort · 18 models · fastest first

Muse Ember23.2 s
09.11.2026 21:17 · 23.2 s · works Works

GPT-5.6 Luna29.7 s
28.10.2026 10:46 · 29.7 s · works Works

GPT-6.1 Terra42.9 s
26.10.2026 20:45 · 42.9 s · works Works

GPT-6 Luna51.5 s
28.10.2026 10:11 · 51.5 s · fails Fails

GPT-5.6 Terra1m 38s
28.10.2026 10:29 · 1m 38s · works Works

Claude Opus 5.51m 42s
30.09.2026 19:51 · 1m 42s · works Works

GPT-5.6 Sol1m 46s
02.10.2026 06:02 · 1m 46s · works Works

Claude Haiku 5.52m 36s
28.10.2026 07:44 · 2m 36s · works Works

Gemini 4 Neon2m 36s
19.11.2026 02:49 · 2m 36s · works Works

Claude Sonnet 5.52m 37s
30.09.2026 16:08 · 2m 37s · works Works

GPT-6 Sol2m 39s
02.10.2026 05:47 · 2m 39s · works Works

Claude Fable 5.12m 50s
28.10.2026 11:10 · 2m 50s · works Works

GPT-6.2 Sol2m 56s
17.11.2026 23:15 · 2m 56s · works Works

GPT-6.1 Sol3m 11s
14.10.2026 07:24 · 3m 11s · works Works

GPT-6.2 Astra3m 54s
17.11.2026 23:19 · 3m 54s · works Works

Inkling 2 Mini4m 07s
21.11.2026 07:24 · 4m 07s · works Works

GPT-6 Astra4m 17s
08.10.2026 19:01 · 4m 17s · works Works

Claude Fable 5.24m 43s
09.11.2026 21:52 · 4m 43s · works Works
Not run at low effort: Qwen3.8 27B.
Medium effort · 18 models · fastest first

Muse Ember32.2 s
09.11.2026 21:19 · 32.2 s · works Works

GPT-6.1 Terra47.9 s
26.10.2026 20:46 · 47.9 s · works Works

GPT-5.6 Luna59.1 s
28.10.2026 10:48 · 59.1 s · works Works

GPT-6 Luna1m 21s
28.10.2026 10:13 · 1m 21s · works Works

GPT-5.6 Terra1m 48s
28.10.2026 10:31 · 1m 48s · works Works

GPT-5.6 Sol2m 20s
02.10.2026 06:05 · 2m 20s · fails Fails

GPT-6 Astra4m 15s
28.10.2026 08:53 · 4m 15s · works Works

Claude Sonnet 5.54m 44s
30.09.2026 16:03 · 4m 44s · works Works

GPT-6.1 Sol4m 52s
02.10.2026 05:23 · 4m 52s · works Works

Claude Fable 5.14m 53s
28.10.2026 11:14 · 4m 53s · works Works

Claude Haiku 5.55m 06s
28.10.2026 07:47 · 5m 06s · works Works

Inkling 2 Mini6m 50s
21.11.2026 07:30 · 6m 50s · works Works

GPT-6.2 Sol7m 02s
17.11.2026 23:23 · 7m 02s · works Works

Claude Fable 5.27m 47s
09.11.2026 21:58 · 7m 47s · works Works

Gemini 4 Neon7m 55s
19.11.2026 03:49 · 7m 55s · works Works

Claude Opus 5.58m 56s
30.09.2026 19:41 · 8m 56s · works Works

GPT-6.2 Astra9m 42s
17.11.2026 23:31 · 9m 42s · works Works

GPT-6 SolNot recorded
Works
Not run at medium effort: Qwen3.8 27B.
High effort · 18 models · fastest first

Muse Ember28.0 s
09.11.2026 21:20 · 28.0 s · works Works

GPT-5.6 Terra1m 00s
28.10.2026 10:34 · 1m 00s · works Works

GPT-5.6 Luna1m 04s
28.10.2026 10:50 · 1m 04s · works Works

GPT-6.1 Terra1m 12s
28.10.2026 07:33 · 1m 12s · works Works

GPT-6 Luna1m 27s
28.10.2026 10:15 · 1m 27s · works Works

GPT-5.6 Sol2m 34s
28.10.2026 08:35 · 2m 34s · works Works

GPT-6 Sol4m 12s
28.10.2026 07:36 · 4m 12s · works Works

GPT-6.1 Sol5m 15s
28.10.2026 07:05 · 5m 15s · works Works

Claude Haiku 5.55m 22s
07.10.2026 21:22 · 5m 22s · works Works

Claude Sonnet 5.56m 45s
28.10.2026 07:54 · 6m 45s · works Works

GPT-6 Astra7m 30s
28.10.2026 08:58 · 7m 30s · works Works

GPT-6.2 Astra9m 43s
17.11.2026 23:54 · 9m 43s · works Works

GPT-6.2 Sol11m 10s
17.11.2026 23:42 · 11m 10s · works Works

Gemini 4 Neon11m 56s
19.11.2026 03:57 · 11m 56s · works Works

Claude Opus 5.513m 48s
28.10.2026 09:25 · 13m 48s · works Works

Claude Fable 5.114m 06s
28.10.2026 11:20 · 14m 06s · works Works

Inkling 2 Mini22m 39s
21.11.2026 07:38 · 22m 39s · works Works

Claude Fable 5.234m 55s
09.11.2026 22:30 · 34m 55s · works Works
Not run at high effort: Qwen3.8 27B.
Extra high effort · 19 models · fastest first

Muse Ember52.1 s
09.11.2026 21:22 · 52.1 s · works Works

GPT-6.1 Terra1m 15s
26.10.2026 20:49 · 1m 15s · works Works

GPT-5.6 Luna1m 25s
28.10.2026 10:52 · 1m 25s · works Works

GPT-5.6 Terra3m 50s
28.10.2026 10:36 · 3m 50s · works Works

GPT-5.6 Sol4m 05s
28.10.2026 08:39 · 4m 05s · fails Fails

GPT-6 Luna4m 24s
28.10.2026 10:18 · 4m 24s · works Works

GPT-6 Sol7m 11s
28.10.2026 08:12 · 7m 11s · works Works

Claude Haiku 5.58m 28s
07.10.2026 21:28 · 8m 28s · works Works

GPT-6.1 Sol11m 37s
28.10.2026 07:42 · 11m 37s · works Works

GPT-6.2 Astra15m 14s
18.11.2026 00:22 · 15m 14s · works Works

GPT-6.2 Sol16m 03s
18.11.2026 00:05 · 16m 03s · works Works

Inkling 2 Mini16m 38s
21.11.2026 08:02 · 16m 38s · works Works

Qwen3.8 27B20m 19s
11.11.2026 06:18 · 20m 19s · works Works

Claude Sonnet 5.521m 18s
28.10.2026 08:01 · 21m 18s · works Works

Gemini 4 Neon22m 26s
19.11.2026 04:10 · 22m 26s · works Works

GPT-6 Astra25m 55s
28.10.2026 09:07 · 25m 55s · works Works

Claude Opus 5.528m 09s
28.10.2026 09:39 · 28m 09s · works Works

Claude Fable 5.132m 58s
28.10.2026 11:35 · 32m 58s · works Works

Claude Fable 5.240m 03s
09.11.2026 21:11 · 40m 03s · works Works
Max effort · 13 models · fastest first

Muse Ember47.7 s
09.11.2026 21:24 · 47.7 s · works Works

GPT-5.6 Luna2m 22s
28.10.2026 10:54 · 2m 22s · works Works

GPT-5.6 Terra4m 04s
28.10.2026 10:41 · 4m 04s · fails Fails

GPT-6 Luna4m 38s
28.10.2026 10:23 · 4m 38s · works Works

GPT-6.1 Terra4m 55s
26.10.2026 21:00 · 4m 55s · works Works

GPT-5.6 Sol7m 53s
28.10.2026 08:44 · 7m 53s · fails Fails

GPT-6 Sol13m 48s
28.10.2026 08:21 · 13m 48s · works Works

GPT-6.1 Sol16m 56s
28.10.2026 07:55 · 16m 56s · works Works

Claude Haiku 5.520m 00s
07.10.2026 21:39 · 20m 00s · works Works

GPT-6 Astra36m 32s
28.10.2026 09:34 · 36m 32s · works Works

Claude Fable 5.136m 52s
28.10.2026 12:09 · 36m 52s · works Works

Claude Opus 5.566m 39s
28.10.2026 14:40 · 66m 39s · works Works

Claude Sonnet 5.587m 03s
28.10.2026 13:12 · 87m 03s · works Works
Not run at max effort: Claude Fable 5.2, GPT-6.2 Astra, GPT-6.2 Sol, Inkling 2 Mini, Qwen3.8 27B, Gemini 4 Neon.
The number is the build time; the models at one effort share one scale.
Also on CmdBench
Image-to-world builds. One picture, four 3D worlds.
Thinking Machines' Inkling 2 Mini, running locally on one GPU, built the island cabin picture as a Three.js world on low, medium, high and extra high effort, from the same prompt and the same copy of Three.js as the other runs. One run each, with no human edits. High and extra high stopped at the 60-minute limit, and extra high's world is blank.
Not a benchmark. These builds have no checks, scores or verdicts.
Also on CmdBench
One-prompt builds
OpenAI's GPT-6.2 Sol and GPT-6.2 Astra, released on 17.11.2026, each built a first-person voxel sandbox game in Codex CLI from the same prompt and no picture: a generated block world to walk through, mine and build in. One run each on low, medium, high and extra high effort, with no human edits.
- 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.











