Architect Says Claude Opus 5.5 Drew a House Set Good Enough for a Permit After Two Redline Rounds
A.I. / news
Architect Says Claude Opus 5.5 Drew a House Set Good Enough for a Permit After Two Redline Rounds
Stephen Gallagher's informal test of 13 models gave Opus 5.5 the most human feedback, and the drawings still missed a required second exit.
An architect says Claude Opus 5.5 produced a construction drawing set for a simple house that he judged good enough for a building permit, after about 12 hours of work and two rounds of his redlines.
Stephen Gallagher, an architect who has worked in the building industry since the 2000s, published the test on Substack on Oct. 28. It was discussed on Hacker News on Nov. 2.
What he ran
Gallagher gave 13 AI models rough concept sketches for a house plus a reference set from a different house drafted in AutoCAD. The models produced drawings by writing Python code, not by driving CAD software.
Results ranged from a refusal by Claude Opus 3 to what he called cartoonish drawings from most 2025 and early 2026 models. He wrote that "None of the models released prior to September was able to draw the floor plans without significant errors." Grok 4.7 was the weakest of the September models.
Only Opus 5.5 drew the dual-pitch roof massing correctly: "Opus 5.5 was the only one that got the roof massing correct." Anthropic released the model on Sept. 22, Thurrott reported, and said external evaluators including Frontier Design and METR tested it before release.
| Item | Reported figure |
|---|---|
| Models tested | 13 |
| Time for the house set (Opus 5.5) | about 12 hours |
| Python in the final code (Opus 5.5) | more than 15,000 lines |
| Redline rounds | 2 |
Gallagher put the compute bill at a few hundred dollars. "Redlining Opus 5.5's drawings felt surprisingly routine," he wrote.
The second test and the faults
In a second test Opus 5.5 drafted plans, elevations and details of a historic building from photographs and field sketches, in the style of Frank Snyder's plates. Gallagher wrote that the output beat the work of many architecture firms but was less polished than Snyder's originals and held subtle errors a skilled architect with experience in that construction type would likely avoid.
The house set also omitted a second means of egress, which a Massachusetts state amendment requires. Gallagher partly blamed his own concept sketch for that. The code was only partly parametric: changing the building length stretched the elevations but not many of the plans.
Why the comparison is uneven
Gallagher says the test is his own informal assessment of a small number of tasks, not a formal benchmark. He also says Opus 5.5 received the most human feedback while weaker models received little or none, so the field was not level.
A missing egress is the kind of fault a plan reviewer checks by rule, and the post reports no review by a permitting office. "Good enough for a permit" is therefore Gallagher's judgement, not a municipality's decision.
Opus 5.5 also featured in our report on 50 GPT agents in Minecraft, and Anthropic's own capability claims are weighed in our piece on its exploit-chain research.
Thurrott reported that Anthropic calls Opus 5.5 its best-performing model and says it costs 40 percent less to operate than Opus 5. Claude Sonnet 5.5 and Haiku 5.5 are due next, according to the same report.
Sources
More in A.I.
- 01Meta Prices Muse Ember at $0.10 per Million Input Tokens, Says Nothing on WeightsMeta's smallest model scores 36 on Artificial Analysis, level with Claude Haiku 5.5, but it trailed Haiku on one of the two coding tests the firm ran.
- 02Anthropic Sets $70 to $78 Range for Nasdaq IPO, Crunchbase ReportsAn amended filing reportedly adds a price range and a $26 billion private placement, and shows a nine-month net loss of $104 billion on $33.2 billion in revenue.
- 03Google's Gemini Help Page Says Human Reviewers See Drive and Gmail ExcerptsA GitHub post from a user named thisisme6598 claims Gemini reads files unprompted. Google's own page confirms review, but not that part.
- 04Anthropic Ships Claude Fable 5.2 at $8 per Million Input TokensArtificial Analysis ranks the model first of 74 on its index, but its run needed 275 million output tokens and 44 minutes per task.