Construction AI: OpenAI ships GPT-5.6 with a price list, and the question of whether your BIM platform can ever hand it the keys

On 26 June 2026 OpenAI previewed GPT-5.6, its new general-purpose frontier family, in three tiers: Sol as the flagship, Terra as the balanced everyday model, and Luna as the cheap, fast option, with a compute-intensive Sol Ultra mode sitting above Sol. The capability story is real but not the interesting bit. The interesting bit is that OpenAI published the prices. Sol runs at US$5 input and US$30 output per million tokens, Terra at US$2.50 and US$15, and Luna at US$1 and US$6 (OpenAI's own figures). For the first time you can read the per-token cost of the new frontier off a page.

So why should a contractor care about a token price? Because it's the number that decides whether an agentic workflow is affordable on an actual job. A copilot that answers a one-line question is cheap on any tier. An agent that loops over a 400-page operation and maintenance pack, reading, checking, cross-referencing, burns output tokens at a rate that turns Sol's thirty-dollar rate into real money fast, while Luna's six does the same job for a fifth of the cost if it's good enough. When you price that work in Intelligence Units, the tier you choose is the line that moves the bill, not the brand on the box. I'd not over-read OpenAI's benchmark either, the TerminalBench 2.1 scores of 88.8% for Sol and 91.9% for Sol Ultra are the vendor's own and measure agentic coding, not quantity surveying. The direction is right, the number needs a real job to test it.

There's a catch you can't ignore. You can't buy it yet. During the preview the models reach only about twenty trusted partner organisations, through the API and Codex, and not in ChatGPT at all, a restriction OpenAI says it took at the US government's request following a 2 June executive order. OpenAI went further than most would, saying publicly it believes in broad access and that this kind of gating shouldn't become the norm. Worth noting that this happened the same week Anthropic's strongest model was being rationed to named critical-infrastructure operators, a thread we covered on 29 June. Two of the biggest labs, the same gate, within days. The pattern is now the story.

The procurement filter: When you cost any AI copilot for a project, ask which model tier it runs on and what that does per Intelligence Unit at the volume you'd actually use, not the demo volume. A five-fold price gap between Sol and Luna is the difference between an agent that pays for itself and one that doesn't.

Picture the gate lifting and GPT-5.6 arriving on your desk tomorrow. Where would you point it? At your BIM model and your CDE, the places your project actually lives. And this is where the release runs into a wall the trade press has been circling for months. The clearest write-up is Martyn Day's piece in AEC Magazine on 28 April, which reads a Google DeepMind paper on AI delegation and concludes, fairly bluntly, that today's BIM platforms are architecturally incompatible with safe agentic delegation. Not incomplete. Incompatible.

The distinction he draws is the one that matters on a high-risk building. There's a difference between an AI that assists, suggesting a layout or drafting a schedule that you then validate and own, and an AI you delegate a decision to, where the agent itself takes responsibility for satisfying the fire egress rules, the structural limits and the accessibility code all at once, and proves it did. Current tools do the first. They can tell you whether a clash was detected. They can't tell you whether the agent honoured the energy target, or how it reached its answer, or sign that answer with something auditable later. The reasoning is a black box, and in structural safety or fire compliance a black box is exactly what you can't have. The comparison only goes so far, but we don't accept opaque reasoning in bridges or aircraft, and a higher-risk building sits in the same bracket.

What makes this a now problem rather than a someday one is that the vendors are already shipping. Autodesk Assistant, Bentley Copilot and Trimble Agent Studio are all agent-capable platforms out in the market, while the draft revision of ISO 19650, the standard that governs how building information is managed and whose Part 3 is open for comment right now, says nothing about agents, autonomous workflows or delegated authority at all. It still assumes a human produces the information and a human is accountable for it. The market is moving a full revision cycle faster than the standard. That gap isn't academic. It's precisely where the liability lands when something goes wrong and nobody can say which layer made the decision.

For your board pack: Before you let any agent-capable platform near a live UK job, ask the vendor two plain questions: can you show me how the agent reached its answer, and whose name is accountable for it under the RICS standard? If the honest answer to the first is no, you've found the limit of what you can safely delegate today.

Tie the two together and you get the honest read on where most UK firms actually are. The model frontier is sprinting, gated and priced. The platform underneath it can't yet be trusted with a delegated decision. And the workforce sits behind both: RICS survey work still puts roughly 45% of construction organisations at no AI use at all, with skills shortages, poor data quality and integration problems named as the brakes. So the gap between what the technology can do and what the average firm can use it for is, if anything, widening this quarter, not closing.

That's not a counsel of despair, it's a steer on where to spend the summer. You're not falling behind because you can't get GPT-5.6, you can't, and nor can almost anyone. You fall behind by leaving your project data and your sign-off chain in a state where no agent, this year's or next year's, could be trusted with them. Clean information, a clear accountable owner for every output, and an honest map of where an agent's reasoning would have to be visible before you'd rely on it. That's unglamorous work and it's the work. Get it right and you're ready the day the gate lifts. Skip it and the cleverest model in the world has nowhere safe to stand.

A practical step: Pick one workflow this month, the O&M handover or the Gateway 2 documentation, and ask of it: if an agent did this, could I prove what it did and name who's responsible? Fixing the gaps that question exposes is worth more than any model you can't buy yet.