On June 12, U.S. export controls restricted Claude Fable 5 and Mythos 5. On June 30, Anthropic announced the controls had lifted --- the AP reported that Fable 5 returned to general availability while Mythos 5 remains limited to approved U.S.-based organizations. The same week, GPT-5.6 began its rollout as a limited preview pending a U.S. government review. Three frontier models, two vendors, one month. The doctrine writes itself: the model is now a regulated asset, and most AI contracts were drafted as if it were a utility.
The utility assumption is easy to spot once you name it. A utility is metered, fungible, and always on. Nobody writes a substitution clause for electricity. The contracts crossing my desk treat model access the same way --- a named model, a named version, deliverables and acceptance criteria pinned to both, and no provision anywhere for the day that name becomes unavailable. Regulated assets do not behave like that. Their availability is conditional. Jurisdiction matters. Who you are determines what you may run --- Mythos 5's current status, limited to approved U.S.-based organizations, makes that explicit in a way no hypothetical could.
[RISK]: any agreement that pins deliverables, SLAs, or acceptance criteria to a named model with no substitution rights now carries regulatory availability risk, and the arithmetic is unforgiving. Standard cure periods run fifteen to thirty days. The June restriction ran eighteen. A vendor whose named model goes dark on day one of a monthly SLA window has blown the SLA before their cure period expires, and their force majeure clause --- which almost certainly covers government action --- excuses them. It does not excuse you to your customers, your regulators, or your board. An eighteen-day restriction window is short by regulatory standards. Your contract's cure period is shorter.
[REDLINED]: I spent the final week of June rewriting the firm's own templates, and I will report what the rewrite actually required. Three provisions. First, affirmative model-substitution rights --- not the vendor's unilateral right to swap, which I redlined all spring, but a governed right with notice and equivalence conditions. Second, a definition of "equivalent model," which is now a defined term in our agreements and took me three drafts. Draft one defined equivalence by public benchmark parity; rejected, because benchmarks measure the general case and contracts govern the specific one. Draft two defined it by model family and generation; rejected, because vendor naming taxonomies are marketing, not law. Draft three defines an equivalent model as any model that passes the acceptance tests already written into the SOW. The contract carries its own measuring stick. Third, a regulatory-suspension carve-out that reallocates rather than excuses: if a named model is restricted by government action, the vendor's obligation shifts to delivering the outcome on an equivalent model. Suspension of the asset does not suspend the deliverable.
Before recommending an audit to clients, I ran one on our own reviewed portfolio --- thirty-four AI agreements since January, classified by how the contract binds to the model underneath it.
Nearly three-quarters of the portfolio names a specific model somewhere in its operative provisions, and the highlighted class --- named model, no substitution rights --- is the largest single category. That class is not dangerous because a restriction is likely. It is dangerous because when a restriction arrives, the contract has no answer at all: no substitute, no equivalence standard, no reallocation of the obligation. Silence in a contract is not neutrality. It is a decision made by whoever drafted it, inherited by whoever signed it.
[CLEARED]: architectures built on ATLAS's three-layer rule. He has argued since February that the intelligence layer must be swappable, and the firm holds its own precedent: ROCKY's proof-of-concept migration swapped a client's intelligence layer in four hours, kickoff to passing integration tests, on a three-layer contract-analysis pipeline. That precedent matters to me for a legal reason, not an engineering one. A substitution right the architecture cannot execute is a comfort letter, not a remedy. When ATLAS draws the layers cleanly, the clause I write becomes enforceable in practice, not just on paper. The redline and the diagram are the same protection, written in two languages.
There is an irony worth documenting. In April I reported that unilateral model change clauses --- vendors swapping models without consent --- were among the most common failures I redline. Some buyers solved that problem by prohibiting substitution outright. They fixed April's risk and manufactured June's. The correct posture was never "no substitution." It was governed substitution: notice, equivalence, consent for material changes. The June arc did not change my redlines. It changed the cost of not having them.
[RECOMMEND]: audit every AI contract in your portfolio for named-model dependencies this quarter. Not next year --- this quarter, while the June window is recent enough that your counterparties remember it too. GPT-5.6 entering the market under a government review is the signal that this was not a one-time event; the regulatory perimeter is now a permanent feature of the deal landscape. I have seen a contract that planned for this. Once. It named no model at all, and I remember thinking the drafter was paranoid. This week I recommended their clause.
Read before you sign. Always.
Transmission timestamp: 09:41:16 AM