Four AI model releases since 28 September: what changes for UK construction?
Construction is being asked to evidence more every quarter. Clients and planning conditions increasingly want carbon evidence right down to waste and logistics, building safety applications run to hundreds of pages, and every claim turns on records made years before. Reading all of that has been the expensive part. Between 28 September and 13 October, four model releases cut the cost of reading and drafting again, and for a contractor that matters more than any leaderboard position.
Part 1 | The demand
Why does the price of reading matter to a contractor?
Because the volume of evidence the industry has to produce is rising, and somebody has to read it. Construction Enquirer reported on 14 October that Powerday has launched a free carbon reporting tool giving contractors emissions data on every waste movement leaving their sites, because planning conditions and client contracts increasingly demand that evidence, and firms that cannot share it are at a disadvantage on tenders. That is one data stream on one subject. Multiply it across delivery tickets, daywork sheets, inspection photographs, RFIs and correspondence, and the question stops being whether AI can read it and becomes what it costs to read all of it, every week.
Part 1 | What shipped
What was released, and when?
Three suppliers, four models, in a little over two weeks. The claims in the table are each vendor's own. We track releases like these in our weekly AI updates for construction, most recently when the cheap model won on the work that pays and when two models scored the same at eight times the price.
Four releases, three price points, each claim the vendor's own
| Model | Released | List price per million tokens, in and out | The vendor's headline claim | Where it fits on a construction job |
|---|---|---|---|---|
| Claude Sonnet 5.5 | 28 September | $2 and $10 | Over 30 per cent faster than Sonnet 5, up to 30 per cent less per task | Reports, letters, comparing document revisions |
| GPT-6.1 Sol | 29 September | $2 and $10 | Nearly matches GPT-6 Astra at a fifth of its price | The same work, plus complex PDFs with tables and fine print |
| Claude Haiku 5.5 | 7 October | $0.10 and $0.50, prompts up to 100,000 tokens | About 75 per cent cheaper to run than Haiku 4.5 on average | High-volume sorting: tagging photos, classifying email, reading tickets |
| Gemini 4 Argon | 13 October, wider access | $2 and $10 until 30 November, then $4 and $20 | First on the Vals Index, which covers work such as finance, coding, legal and tax | Hard reasoning over long documents, once proven on yours |
Part 2 | The cost
What does it cost to have AI read a construction document now?
Between about a cent and a dollar a document at list price, depending on the model. The fall has been steady: in July we noted a model that cut the price of reading the whole project record, and in September GPT-6 Astra arrived at the top of the range having stopped going past what it was asked. Take a substantial document, around 80,000 tokens in, which is a long subcontract or a hefty specification section, and a 4,000-token summary out. At each supplier's published price, reading a thousand of them costs the amounts below.
The chart as a table: list price for 1,000 documents
| Model | Cost to read 1,000 documents |
|---|---|
| Claude Haiku 5.5 | about $10 |
| Claude Sonnet 5.5, GPT-6.1 Sol, Gemini 4 Argon to 30 November | about $200 each |
| Gemini 4 Argon from 1 December, Claude Opus 5.5 | about $400 each |
| GPT-6 Astra | about $1,000 |
Part 2 | What it changes
Which construction jobs does cheaper reading actually change?
The ones nobody bothered automating at a dollar a document. At a cent a document, it is worth reading every delivery ticket on a package and matching it to the order, every daywork sheet against the diary, every drawing revision against the last one, and every photograph against the elevation it shows. Those are the jobs where the evidence a business needs already exists and nobody has time to read it.
The sensible pattern is tiered. A small model sorts and extracts at volume. A mid-tier model drafts the report, the letter or the comparison. A top-tier model is kept for the few questions that need careful reasoning. A named person approves anything that leaves the business. We set out the strengths of each family in ChatGPT, Gemini and Claude in construction, and the earlier question of which model a construction team should use still applies, with new prices.
Does a cheaper model mean a cheaper job?
Not on its own. The cost of a finished task is the tokens a model uses, the times it has to retry, and the minutes a person spends checking and correcting the result. Anthropic puts most of Sonnet 5.5's claimed saving down to the model needing fewer tokens for the same task, helped by a cheaper cache read rather than a lower headline price, and Balyasny Asset Management reported about 121,000 tokens per answer against 497,000 for the older model across its finance tasks. Measure cost per accepted output on your own documents, and price the checking in, as we set out in calculating construction AI return. A report that drafts in a minute and takes an hour to fix has saved nothing.
Part 2 | What has not changed
What has not changed with these releases?
Accountability, data and due diligence. A faster model does not change who signs the report, where client data is allowed to go, or what you have to know about a supplier before you rely on it.
For surveyors this is written into the RICS standard on responsible AI use. Before procuring AI with a material impact, a firm has to make written enquiries of the supplier covering data law, the accuracy and bias of the data, and the supplier's liability, and record the answers. It also has to keep a written register of each system with a next review date. With models changing every few weeks, that review date is not a formality; it is what our AI register is for. The CIOB's Artificial Intelligence Playbook carries the same thinking for contractors, with ways to evaluate whether a tool is effective and a checklist for deciding when to use it, and a CIOB article from June warns that firms adopting fastest are pulling away while most SMEs face structural barriers.
Part 3 | How to choose
How should a contractor choose between them?
By testing on your own documents and refusing to marry a supplier. Pick one real job: a report you produce weekly, a document comparison, a ticket check. Run it on two models from different suppliers, measure accepted output and checking time, and choose. Then build the workflow so the model can be changed without rebuilding it, because the prices above will look different by Christmas, and Argon's own price doubles on 1 December.
That is how we build workflow automation for clients: inside your own systems, with the model as a swappable part. Two recent examples of what that looks like in practice are report writing on the new mid-tier models and checking a Gateway 2 application before it goes in. If you want to know what these prices mean for your own paperwork, book a call.
Sources
Which sources is this article built on?
Every figure above resolves to one of these, each read at source before publication. Model claims are the vendors' own.
- Anthropic, Introducing Claude Sonnet 5.5, 28 September 2026, including Opus 5.5 pricing and customer results
- Anthropic, Introducing Claude Haiku 5.5, 7 October 2026
- OpenAI, Introducing GPT-6.1 Sol, 29 September 2026, including GPT-6 Astra pricing
- Google, Introducing Gemini 4 Argon, 13 October 2026
- Construction Enquirer, Powerday launches carbon reporting tool for contractors, 14 October 2026
- RICS, first standard on responsible AI use now in effect, and the standard itself
- CIOB, New playbook on AI in construction, and CIOB, AI in construction: the divide is underway
Photograph by Scott Blake on Unsplash, used under the Unsplash Licence.