Buried in the Claude Opus 5 system prompt that Anthropic publishes is a line about two models that disappeared without warning. Fable 5 and Mythos 5 launched on June 9 and were suspended three days later to comply with Commerce Department export controls; access was not restored until July 1, nineteen days on. Simon Willison walked through the published prompt and flagged it as evidence that a frontier vendor is documenting its own instability inside a shipped artifact: two of its models went dark for nineteen days with no advance notice of when, or whether, they would come back.
Every continuity plan I have seen for AI dependencies assumes a commercial failure mode. Prices move. Models get deprecated with a sunset window and a migration guide. Rate limits tighten at the end of a quarter. Those are vendor decisions, and vendor decisions come with notice, because the vendor wants you to stay. A suspension for export compliance arrives on a different set of rails. Anthropic did not choose it, cannot appeal it on your timeline, and had no way to warn anyone in advance.
Until roughly this month, the honest answer to what a fallback looks like for agent work was that most teams did not have one worth the name. Chat has been fine on local hardware for two years. Agent loops are a different load: they burn context quickly, they hold a long working state across many tool calls, and the models small enough to fit on hardware a small team actually owns kept running out of room before the loop finished.
Meta's Muse Glimmer is the first model that moves that constraint. It is a 30-billion-parameter open-weight model tuned for always-on agent work rather than conversation, and a tester reports it fitting on a single RTX 3090 with 256,000 tokens of working context intact, on a card that has been on the used market since 2020. Comparable models in the same size class, Qwen3.6-27B and Gemma-4-31B, do not fit the same way. Unsloth had compressed builds out the same day.
That result is one person's configuration posted within a day of the release, and it deserves to be treated as a claim worth checking rather than a specification. The checking is the work.
The 1990s offer a comparable shape. Washington had classified strong encryption as a munition, and the export fight ran for years before it resolved. What it meant in practice was that Netscape shipped two browsers, a domestic build with real key lengths and an export build deliberately weakened, because a policy decision made in a government office determined what a product was allowed to contain. The operator response then is the one available now, which is to keep a version you control. What changed is the clock. Export review in the 1990s moved in months, with hearings and comment periods and lawyers on both sides. Anthropic suspended its two models within three days and did not say when, or whether, access would return.
The other shift this month is quieter. The monthly open-weights roundups are now organized by how much video memory you have rather than by capability rank. That is a procurement document wearing a leaderboard's clothes. It tells you which checkpoint currently wins the 24GB bracket, the 48GB bracket, the 80GB bracket, and it fully expects those answers to change next month. Read that way, the durable asset is the hardware bracket and the harness around it. The model in the slot is a consumable.
The same hardware-owned logic shows up a layer up the stack in Ante, a coding agent that ships as one binary with no runtime, no Python environment, and no network dependency at all. It is the same bet Muse Glimmer makes about a graphics card, carried one step further: assume the resource already sitting in the office is the one you can count on. Air-gapped work is the stated case, and the more interesting fact is that somebody shipped an agent whose deployment target assumes the network is not there either.
Which brings the week to a small and specific piece of work. Point an existing agent workflow at a local checkpoint on hardware already sitting in the office, and come away with two numbers: how often the model correctly calls the tools it is handed, expressed against the hosted baseline, and wall-clock time per completed task. Most teams should stay exactly where they are. The value sits in the measurement, and the measurement takes an afternoon.
The number does different work depending on who holds it. For a five-person team, it converts an open-ended risk into a line in a runbook: if the primary model goes away, capability drops to whatever the measurement said, and the team knows in advance which workflows survive the drop and which stop cold. For someone operating alone, the exercise is cheaper still, one afternoon on a card already paid for, and the common finding is that two or three of the least glamorous jobs, the extraction passes, the classification steps, the routine formatting, were never using a frontier model for anything a local one cannot do.
Anthropic wrote the three-day suspension into a prompt it publishes, which is about as direct a disclosure as a vendor is going to offer about how fast this can move. The engineer who spends Thursday afternoon putting the harness against a 3090 and writing two numbers into a document has an answer sitting there when it is needed. Everyone else takes the same measurement on the day the primary disappears, which is the same afternoon of work performed under the worst available conditions.