Construction AI: Anthropic accuses Alibaba of scraping Claude at scale, and the hidden cost tucked inside Sonnet 5's price tag

Here's the headline, and then the bit that actually matters to a construction business. On 10 June 2026 Anthropic wrote to the US Senate Commerce Committee, to chair Tim Scott and ranking member Elizabeth Warren, alleging that operators tied to Alibaba's Qwen lab opened roughly 25,000 fraudulent accounts and pushed about 28.8 million exchanges through Claude between 22 April and 5 June. The letter surfaced publicly around 24 June. The technique it describes is distillation: point your cheaper model at a stronger one, harvest millions of its answers, and train yours to imitate them. Anthropic says the exchanges deliberately targeted software engineering and agentic reasoning, which happen to be the most commercially valuable things Claude does. It's the first time Anthropic has named a major Chinese technology conglomerate, and by its own account this single campaign is bigger than its February accusations against DeepSeek, Moonshot AI and MiniMax put together. Alibaba denies the lot. I'll say it plainly: this is an allegation, not a finding, and it's slightly older than my usual four-day window, but it's significant enough to carry.

So why does a quantity surveyor or a BIM lead need to know any of this? Because of a thread we keep pulling on. The big move in AEC through 2026 has been "local AI", the idea that you run an open-weights model on your own hardware, or a sovereign cloud, so your golden-thread data and your commercially sensitive project files never leave your control. It's a genuinely good instinct, and it's the right answer to a lot of the data-governance worry in this sector. But the strongest open-weights models you'd actually reach for, the ones that top the open leaderboards, are largely Chinese: Qwen, DeepSeek, Kimi, GLM. And Qwen is now the one standing accused of having built itself by quietly copying a Western rival. What that does is turn "which local model do we run" from a pure benchmarks-and-cost question into a provenance question too. The comparison only goes so far, but it's a bit like specifying a cladding product that's brilliant on paper while an investigation into how it was tested rumbles on. You might still choose it. You wouldn't choose it without writing down that you knew.

None of this means abandon local AI, and it certainly doesn't mean the Chinese models are bad tools, they're often excellent. It means the due diligence a serious firm does before it standardises on a model now has an extra line in it. Who made this, how, and is there a live legal cloud over it that a client or an insurer might one day ask me about. That's a five-minute conversation to have now and an awkward one to have later.

The procurement filter: Before you standardise on any local open-weights model, write one paragraph on its provenance and any live IP disputes, and put it next to the cost and benchmark case. If you can't, you're not ready to make the choice yet.

On Monday I told you Sonnet 5 arrived cheap and ungated, and that the last easy excuse for not using a capable agent had gone. Both still true. But a detail I didn't have then changes the maths, so here it is. Sonnet 5 ships with a new tokenizer, the thing that chops your text into the billable units a model reads. Anthropic's own documentation confirms it produces roughly 30% more tokens for the same text than Sonnet 4.6 did, and independent testers who've run it back that up. It's not uniform: about 27% more for code, and up to something like 42% more for ordinary English prose, which is most of what a construction workflow feeds a model, think RFIs, O&M narratives, meeting notes, safety-case text.

The per-token price hasn't changed, which is exactly why this is easy to miss. Anthropic held the list rate at US$3 in and US$15 out, with the introductory US$2 / US$10 running until 31 August. But if the same document now counts as a third more tokens, then at the standard rate the same job costs you roughly a third more than it did on the previous model, for identical work. The intro discount papers over that until the end of August. Then two things land at once: the rate steps back up to US$3 / US$15, and the token inflation is still there underneath. Independent cost analysts reckon the effective jump for equivalent traffic ends up somewhere in the 20 to 35% range once the discount ends. Treat those specific figures as commentary rather than gospel, they depend heavily on your content mix, but the direction is not in doubt and the mechanism is confirmed by Anthropic itself.

What this means on the ground is unglamorous and important. If you're pricing an AI copilot into a bid, or setting an internal Intelligence Unit allowance for a team, the headline sticker will quietly understate the real cost, and it'll understate it most for the prose-heavy documents construction runs on. Model the actual token count on a real document you already have, and model the September step. The person who gets caught out otherwise is whoever put their name to the budget.

For your board pack: Re-run your Sonnet cost model on a genuine project document, not a sample, and show the number both before and after 1 September. One slide, two figures, no surprises later.

Put the two together and they're really the same instruction wearing different clothes. One's about where your model came from, the other's about what it actually costs to run. Both are questions you answer once, quietly, before you commit, and both are cheap to answer now and expensive to answer after you've standardised on the wrong thing.

The wider evidence says most firms still skip this step. The RICS Construction Productivity Report earlier this year found more than a fifth of UK firms don't measure productivity at all, which means they've no baseline to prove a saving against and no way to notice a cost creeping up on them. So before the next model, the next agent, the next procurement paper, do the boring pair of checks. Where did this come from, and what does a real day's work on it actually cost. Get those two right and most of the rest looks after itself.

A practical step: Pick one workflow you're tempted to hand to an agent this month. Write down its current cost and cycle time first. That single before-number is what turns a pilot into something you can defend.