Matthew Berman put Sonnet 5.5 to work. What does it mean for construction reports?
A routine inspection report rarely takes a day because of the thinking. It takes a day because of the assembly: two hundred photographs to sort and label, a page of notes to type up, the same client headers and standing data to fill in again, and every observation tied back to a drawing and a location. That assembly is a long chain of small, checkable steps, and holding a long chain together is exactly what the newest AI models got better at this autumn. The week Claude Sonnet 5.5 launched showed it clearly, even though nobody was writing a report.
Part 1 | The model
What did Matthew Berman actually do with Sonnet 5.5?
He built things with it and put them online. In his video, published within a day of Anthropic launching the model, Matthew Berman set Sonnet 5.5 to build interactive 3D scenes and browser games, among them an ocean scene with a live settings panel. He published five finished projects and two versions of the same ideas built with the older Sonnet 5, so anyone can open both and compare.
None of that is a construction report, and it would be silly to pretend otherwise. The reason it matters is what a working game demands: hundreds of decisions made in order, rules, assets and controls that all have to agree with each other, and a finish line the model reaches without stopping halfway to ask what to do next. Swap the assets for photographs and the rules for a client's report template and you have described a reporting job.
Part 1 | The claims
What did Anthropic and its early testers claim?
Faster, cheaper per task, the same list price, and better with documents. Every line below is a vendor or customer claim published on launch day, not an independent construction test.
The launch claims that matter to a report writer
| Claim | Who said it | Why it matters for reports |
|---|---|---|
| Output more than 30 per cent faster than Sonnet 5 | Anthropic | A draft comes back while the surveyor is still on site |
| Up to 30 per cent less per task, still $2 and $10 per million tokens | Anthropic | Anthropic puts it mainly down to fewer tokens, so test it on your own files |
| Goes back to the source documents to check figures, and found mistakes Sonnet 5 let through | Box, an early tester | The behaviour you want before a number reaches a client |
| About 121,000 tokens per answer against 497,000, across 2,441 finance tasks | Balyasny Asset Management | Cost per document falls even at the same list price |
| A 10-slide operating review from earnings papers, judged ready to send as drafted by two experts | Anthropic's own test | The first draft is now close enough that checking becomes the job |
OpenAI answered within a day. At its developer conference on 29 September it released GPT-6.1 Sol at the same $2 and $10, claiming results close to its flagship on professional work, including questions about complex PDFs with tables, charts and fine print. Two suppliers competing at that price point is good news for anyone paying to have reports read and drafted, and it follows the summer's move to sell speed as a line on the rate card.
Part 2 | Where the time goes
Where does the time go in a construction report?
Mostly into putting it together. Take a routine condition or progress inspection on a mid-sized commercial building: a couple of hundred photographs, a page of handwritten notes, a voice memo recorded in the car, and a template the client insists on. The surveyor's judgement on what the defects mean and what should happen next is the valuable hour. The other hours are sorting, typing, numbering, formatting and cross-referencing.
The same model as a table: the platform assembles, the surveyor decides
| Stage | Before | After | What the platform does | Who checks |
|---|---|---|---|---|
| Sorting and labelling photographs | 2.0h | 0.3h | Groups by location, elevation and observation | Surveyor, at a glance |
| Typing up notes and voice memos | 1.5h | 0.2h | Transcribes and files each note against its observation | Surveyor |
| Template, standing data, formatting | 1.5h | 0.1h | Fills headers, addresses and job data from the record | Nobody needs to |
| Cross-referencing to drawings | 1.0h | 0.3h | Links each observation to a drawing and grid reference | Surveyor |
| Drafting the observations | 2.0h | 0.7h | First wording in the house style, tied to its photo | Surveyor, line by line |
| Judgement and conclusions | 1.5h | 1.5h | Nothing | The surveyor writes it |
| Checking and sign-off | 0.5h | 0.9h | Flags gaps: no photo, no location, no priority | Named signatory |
| Total | 10.0h | 4.0h | Net of the extra checking | 60 per cent less |
Part 2 | The 60 per cent
How does a reporting platform take 60 per cent out?
By doing the assembly and leaving the judgement alone. Photographs are grouped by location from their metadata and the inspector's notes. Voice memos are transcribed and filed against the observation they describe. Headers and standing data come from the job record, not from retyping. Each observation gets a first wording in the client's style, with its photograph and drawing reference attached, and a completeness check stops anything going out without a location, a photo, a priority and an action. The surveyor then reads every line, corrects what is wrong and writes the conclusions.
That is the design behind our inspection reporting work and our document automation projects. A model like Sonnet 5.5 does the drafting and the reading. The platform around it does the rest: the template, the record, the checks and the audit trail.
Part 2 | Keeping it honest
Is a 60 per cent saving believable?
On the right report, yes, as a design target. Across a business, no, and the arithmetic shows why. If routine reports take a fifth of someone's week and the workflow suits seven reports in ten, a 60 per cent cut on each report is about 8 per cent of the role. That lines up with the evidence on what automated reporting actually saves, which found no end-to-end trial of construction reporting and treats single-digit gains across a role as the honest planning figure.
Both numbers are true. Sixty per cent is what the platform is built to remove from a routine report; 8 per cent is roughly what the business feels. Quoting the first as if it were the second is how AI projects lose credibility. Claims of documents in minutes rather than days, like the one Construction Enquirer reported for an AI platform launched for SME contractors, may well be true of the drafting. What matters is the time once someone competent has read the result.
Part 2 | The limits
What must stay with a person?
The conclusions, the cause of a defect, the priority, anything touching safety and anything with contractual weight. For surveyors this is now written down. The RICS standard on responsible use of AI, in force since 9 March 2026, requires a written reliability decision on material AI output by or under a named, qualified surveyor, randomised dip samples of high-volume automated output, and clients told in writing before AI is used. The CIOB's Artificial Intelligence Playbook makes the same point from the contractor's side: people move from operating the tools to supervising and checking what they produce.
The working rules sit in our human in the loop matrix, and why a fluent first draft is still not a source is covered in hallucination risk in professional documents.
Part 3 | Where to start
How should a firm test this on its own reports?
Pick one report type you produce every week. Time ten of them as they are done today, from site notes to signed issue. Put the workflow in place, run ten more, and count the correction and checking time honestly. If the saving holds, extend it to the next report type. If it does not, you found out for the price of one workflow.
That is the shape of our first projects: one workflow, measured before and after, live in about six weeks. If you want to see what it would look like on your reports, book a call or email chris@ai-metric.com.
Sources
Which sources is this article built on?
Every figure above resolves to one of these, each read at source before publication. The model's claims are the vendor's and its customers', not ours.
- Anthropic, Introducing Claude Sonnet 5.5, 28 September 2026
- Matthew Berman, How Did Anthropic Pull This Off?, YouTube, published 28 September 2026 Pacific time, 29 September in the UK. We followed the model's launch in this video and credit it for the pointer
- OpenAI, Introducing GPT-6.1 Sol, and OpenAI, DevDay 2026 recap, 29 September 2026
- RICS, first standard on responsible AI use now in effect, and the standard itself
- CIOB, New playbook on AI in construction published by CIOB
- Construction Enquirer, AI project management platform launched for SME contractors, 13 March 2026
Photograph by Sven Mieke on Unsplash, used under the Unsplash Licence.