Construction AI: the engine under your software just got 80 per cent cheaper, and the useful stuff is a sensor in the slab

Here's a number that tells you more about your software bill than any product demo. On 30 July OpenAI cut the price of one of its GPT-5.6 models, the cheap tier called Luna, by 80 per cent, from 1 dollar in and 6 dollars out per million tokens down to 20 cents in and 1.20 out. It trimmed the mid tier, Terra, by 20 per cent, and left the flagship, Sol, untouched at 5 and 30. The thing that should make you sit up is the timing. GPT-5.6 only launched on 9 July. So the price came down by four fifths inside three weeks of going on sale.

Why does a company drop its own prices that fast? Because it's being made to. CNBC reported that Chinese open-weight models had climbed to roughly 46 per cent of US enterprise token usage on OpenRouter, a big model marketplace, and were at times running ahead of the American models. Forbes called it a race to the bottom, and I think that's about right from the labs' side of the table. When DeepSeek and Kimi and the rest will do a capable job for pennies, the premium labs either match them or watch the work walk. The pricing power that everyone assumed OpenAI and its peers had is turning out to be thinner than the valuations suggest.

Now the bit for us, because this isn't an AI-industry story, it's a procurement one. Every piece of construction software you rent that has AI in it, the report writer, the document checker, the takeoff tool, the RFI chaser, runs on one of these engines underneath. That engine is now a commodity, and its price is falling by the month. What that means on the renewal you sign this autumn is a fair question the vendor won't volunteer: their cost of running your AI features has dropped hard this year, so has your price followed, or has the saving quietly become their margin? A challenger with published pricing can pass a falling cost straight through to you. A platform that prices per user or per project spend has less reason to. The comparison only goes so far, but it's worth putting the question on the table.

And there's a second, blunter lesson. Don't build your business on one lab's release, or one lab's price. When the whole market can re-rate itself 80 per cent in three weeks, the tool you want is the one that can swap the model underneath without you noticing. Marry a single engine and you've tied your firm to somebody else's pricing decision. That lands on the person doing the takeoff, next renewal, wondering why the bill moved.

While the labs squabble over token prices, look at where a chunk of real construction money went at the end of July: into a lump of electronics you drown in wet concrete. GreenVibe, an Israeli firm, raised 12 million dollars in seed funding, backed by Insight Partners alongside contractors Tidhar, Israel-Canada and Shikun and Binui. Trade press picked it up through the end of the month. What it makes is a wireless sensor you embed in the pour that reports the concrete's strength and structural status every 30 minutes, during the pour and after it, to a phone or a desktop.

Here's why that's more interesting than it sounds. The way most sites check concrete strength hasn't really changed in decades. You take cylinder samples, send them to a lab, and wait days for a result that tells you what the concrete was like in a mould in a lab, not what's actually curing in your slab. GreenVibe's pitch, and flag it as theirs because the figures are the company's own, is up to 95 per cent correlation with the concrete that's genuinely in the structure, against a cylinder test it says carries a 10 to 50 per cent error. It reckons it's already embedded across more than 20 projects and labs, including Beyond by Tidhar, the tallest building in Israel.

What that does on site is give the site manager a real answer to the question that governs the programme: is it ready. Ready to strike the formwork, ready to stress the tendons, ready to load the floor above. Get that call right from real data and you take days out of a concrete frame safely. Get it wrong the traditional way and you either wait longer than you need to, burning programme, or you move too early and carry a risk nobody can see. I'm not going to pretend a sensor in a slab is glamorous, and it's a narrow tool, one job, not a platform. But it's exactly the kind of narrow job that pays for itself, because it hands a decision back to the person on the deck with something better than a hunch and a calendar.

That's the contrast with the item above, really. The clever general-purpose engine is getting cheap and interchangeable. The value is moving to the specific, physical, unglamorous thing that finishes a real task on a real site. Worth remembering next time someone tries to sell you intelligence when what you needed was an answer.

Put the two together and you get a decent rule for the rest of the year. The raw AI is becoming a commodity, cheap and swappable, with the big labs cutting prices inside weeks to hold their ground. The thing worth paying for is the narrow tool that finishes a job your team was going to sit and wait on, whether that's a slab strength reading or a report that writes itself off your site data. So stop buying cleverness and start buying finished work. Ask a vendor to show you the task their tool completes end to end, not the model it runs on, because the model is the bit that's racing to nearly free.

And keep the discipline underneath it. A cheaper engine doesn't fix messy data, unclear ownership of records, or a workflow held together by three spreadsheets and a WhatsApp group. The AI has only ever been as good as what you feed it, and that hasn't changed just because the price of the AI fell. What has changed is that there's now less excuse for a vendor to charge you a premium for the raw intelligence when the raw intelligence is the cheapest part of the stack.

The renewal filter: before you sign anything this autumn, ask the vendor two things. Which model does this run on, and can you change it. And has the price of my licence moved this year in step with the cost of the AI behind it. The answers tell you whether you're buying a tool or renting a margin.