Construction AI: the agents that do your admin got cheaper this week, and the sharpest cyber models went behind a locked door

So, the first week of September gave us four frontier model launches inside 72 hours, and the interesting part isn't which one won a leaderboard. It's what happened to the price of running one.

On 1 September, Anthropic shipped Claude Fable 5.1 and cut the cost of cached-input reads by 75 per cent, from $1.00 to $0.25 per million tokens, while leaving base input and output pricing untouched at $10 and $50 (VentureBeat and MarkTechPost, both 1 September). The day after, Google put out Gemini 3.8 Flash, its third Flash model in six weeks, at exactly the same $0.75 and $3.75 per million tokens as the version three weeks before it (Google's own launch blog, 2 September). Two labs, same move: hold the headline price, drop the cost of the repetitive work underneath.

That repetitive work is the bit that matters on a project. When Fable 5.1's agentic scores roughly doubled, the jump wasn't in the flashy reasoning numbers, it was in long-running, tool-using tasks like Terminal-Bench-Science, which went from 24.7 per cent to 52.6 per cent. Read that in plain terms and it's the drudge, not the genius: reading a full drawing set, cross-checking a revision against the last issue, chasing an RFI that's gone cold. What a cheaper cache read does is make it cheaper for the tool on your desk to re-read the same job again and again without the bill climbing. The comparison only goes so far, but it's a bit like your quantity surveyor suddenly costing a quarter as much every time they open the same file for the fifth time.

There's a catch worth reading before you sign anything, though. Google's Flash price is an introductory rate, and the small print says it doubles on 1 January 2027. So the cheap number you're quoted this autumn has an end date printed on it. The procurement filter: when your AI-first tool comes up for an annual deal, ask the vendor straight what happens to your price when their model provider's introductory window closes, and get the answer in writing rather than in a demo.

Here's the part that didn't make the headlines. Both labs paired their general release with a cyber-capable model, and both put it behind a locked door.

Anthropic's Mythos 5.1 is the same underlying model as Fable 5.1 with the safeguards loosened for cyber defence and life-sciences work, and access is gated to a vetted set of organisations through trusted-access programmes run with the US government, US organisations only (Unite.AI and the Anthropic system card, 1 September). Google did the parallel thing with Gemini 3.8 Flash Cyber, handed to trusted defenders through its Fairwind Programme, which Google's own security team says wrote 2.6 times more correct Chrome vulnerability patches than rival models (Google, 2 September). Treat that 2.6x as a vendor's own number until someone independent runs the test, but the direction is clear enough.

What that means for a UK construction firm is a bit awkward. The most capable tool for defending the systems that hold your project information, your golden thread, your commercial data, currently won't answer a British phone. You're not shut out of good security, and the defensive gains will filter through to the products you already use. But the frontier of it is gated by country, and I don't think that changes soon. For your board pack: stop asking "are we using AI" and start asking "which model reads our project data, where does it run, and who vetted it". That's a governance question a main contractor's client will start asking at pre-qual, and you want the answer ready before they do.

The plainer point sits underneath both models. The same capability that patches a vulnerability can find one, which is exactly why these got gated. On a construction system that now has to hold a defensible record from Gateway 2 through to handover, knowing who and what can touch that record is no longer an IT footnote. It's part of the safety case.

Put the week together and it's two facts pulling in different directions. The cost of the machine doing your admin fell again. The cost of getting the safety-critical version of it stayed gated to a chosen few. Neither of those changes what Building Control asks to see when your scheme reaches the gate.

That's the discipline that survives every model launch. The tool getting cheaper is welcome, and it's a fair reason to expect more work out of the AI-first platform you already pay for. It is not a reason to let that platform quietly widen what it touches, or to stop knowing where your data goes when it does. The models will keep getting cheaper and the labs will keep printing repricing dates in the small print. The record you hand over still has to stand on its own.

Today's action: before your next software renewal, write down two things on one side of paper, what the tool costs to run today and what it's allowed to read. Then ask the vendor to confirm both in writing. That's the whole governance conversation, and it fits on a Post-it.