How well can AI manage a construction job?
We hand the models a project folder and six construction management jobs, from the tender estimate to the month-end cost report, and score how well they do them.
612
Documents
7
Models
65.9
Best average
Results
| # | Model | Run in | Average | Details |
|---|---|---|---|---|
| 1 | Claude Opus 5.5 Anthropic | Claude Cowork | 65.9 | |
Claude Opus 5.5 in Claude Cowork
| ||||
| 2= | Claude Fable 5.1 Anthropic | Claude Cowork | 63.4 | |
Claude Fable 5.1 in Claude Cowork
| ||||
| 2= | Claude Sonnet 5.5 Anthropic | Claude Cowork | 63.2 | |
Claude Sonnet 5.5 in Claude Cowork
| ||||
| 4 | GPT-6.1 Sol OpenAI | ChatGPT Work | 60.5 | |
GPT-6.1 Sol in ChatGPT Work
| ||||
| 5 | GPT-6 Astra OpenAI | ChatGPT Work | 56.8 | |
GPT-6 Astra in ChatGPT Work
| ||||
| 6 | Gemini 4 Argon Google | Gemini app | 44.1 | |
Gemini 4 Argon in Gemini appRun in the Gemini app on a Pro plan. A first try through the Gemini API cut out before it finished, so it isn't shown.
| ||||
| 7 | Claude Haiku 5.5 Anthropic | Claude Cowork | 35.2 | |
Claude Haiku 5.5 in Claude Cowork
| ||||
One run per job, so a few points either way is noise.
The six jobs
- Estimate 31 docs
- Programme 54 docs
- Progress update 92 docs
- Procurement 68 docs
- Contract admin 172 docs
- Cost report 195 docs