Construction AI: the model your agent runs on just got cheaper, and you might not know which one answered

Here's the release that actually mattered this week, and it wasn't a construction product at all. On 22 September Anthropic shipped Claude Opus 5.5, which it called the first model in its new 5.5 family. The headline for a buyer is price: input dropped to $4 per million tokens and output to $20, down from $5 and $25 on Opus 5, so roughly a fifth cheaper for the same work. Anthropic and the trade press also point at stronger performance on the agent side, with VentureBeat reporting it matches the pricier Fable 5.1 on key agentic coding benchmarks at a much lower cost, and a Terminal-Bench 4.0 score of 66.4 per cent against 55.8 (those benchmark figures are vendor and press-reported, so take them as direction, not gospel).

Why does a model release belong in a construction brief. Because the agents doing drawing checks, compliance review and the RFI chasing all run on models like this one, and most of them turn up inside tools you already pay for rather than as something you buy directly. When the model underneath gets cheaper and quicker, the tool gets cheaper to run, and over time that shows up in what the AI-first products can offer at a published price. It's plumbing, but it's the plumbing under a lot of what's landing on your desk.

Now the bit an editor might cut, because it's the interesting bit. The New Stack reported that Opus 5.5 carries safety classifiers that can quietly reroute a flagged request to an older model, cyber-flavoured prompts to Opus 4.8, biology and frontier ones to Opus 5, and it does this mid-conversation without telling the user. So a request you address to 5.5 might, in fact, be answered by 4.8 or 5. That's reported from Anthropic's own documentation rather than a leak, and for most day-to-day use it changes nothing. But think about it on a live job. If you've got an agent running a sequence of checks and you're relying on them being done consistently, "which model actually answered" stops being a trivia question. It's a bit like sending the same query to two surveyors and quietly getting a third to answer one of them. Usually fine. Occasionally the thing that bites you.

So the practical bit is dull and it's worth doing anyway. If your firm is starting to lean on agents, write down which model is meant to do which task, and make "which model answered" something you can go back and check rather than something you take on trust. The person who'll thank you is whoever has to explain, a year from now, why two checks on the same building came out differently.

Add up the construction-technology money that moved in the week to 21 September and it's a healthy number, about $378m across the deals Bricks and Bytes could verify. But look at where the biggest cheque went and it tells you something about this whole cycle. It didn't go to a clever piece of site software or a robot laying blocks. The largest single raise in the tally, announced on 10 September and picked up in this week's round-ups, was TAR, a firm that took $120m at a $1bn valuation, led by Spark Capital, to build off-grid power for AI data centres, reportedly able to stand a system up in about six months. Latitude Media's line on it was blunt: AI data centres can't wait for the grid.

Set that next to the construction-software raises in the same tally, Buildots on progress capture and Adaptive on back-office accounting, both of which we've covered, and the shape is clear. The money going into our world is chasing the unglamorous end of the job, the paperwork and, above all, the power. And the power bet is the one I'd sit with, because it's a direct read-through to a UK story we ran on Tuesday: Nscale's site at Loughton, £2bn and Microsoft behind it, reportedly looking at a fuel-cell maker because a 90MW grid connection won't be ready. When a $1bn startup and a hyperscaler independently reach the same conclusion, that you build your own power rather than queue for the mains, you're watching the market price in a constraint the rest of us keep treating as a detail.

What that means if you're bidding or subcontracting into data-centre work is practical. The developer's power strategy has moved from a background assumption to a live risk on your programme. Ask where the electrons are coming from, ask to see the plan if the answer is "our own generation," and put that date on the risk register above the planning date. The concrete follows the connection, not the consent.

The procurement filter: on any data-centre enquiry this quarter, add one line to your qualification, "confirm the power source and connection date," and treat a vague answer as a red flag, not a detail to sort later.