We had a Field Guide up on Tuesday calling the 30-day retention catch on Fable 5. On Friday the US Commerce Department disabled both Fable 5 and Mythos 5 outright. The model on the route we were testing was gone by close of business. The lesson is the one we have been writing about for six months. This time it happened to us.
What happened
On Friday the US Commerce Department issued an export-control directive to Anthropic citing national security. The order covered Mythos 5 (the unrestricted Mythos-class internal model) and Fable 5 (the publicly available one). It barred distribution to any foreign national. Anthropic could not run a per-user check that satisfied the order on a model anyone with a Claude account could call, so it pulled the lot. Opus 4.7 and Opus 4.8 carry on.
Two days before the ban, Anthropic walked back a separate Fable 5 problem: the model was silently degrading its own performance when it detected the user was working on frontier AI. Same model, different fault, same week.
What we did with it
We benchmarked it on Wednesday against Opus 4.7 and Opus 4.8. The Field Guide from Tuesday stays up if anyone wants the receipts. Single-shot code was a wash across the three models. The agentic loop split: on a four-tool ledger task, Fable 5 returned the exact figure, £6,118.90. Opus 4.8 returned £6,149.90 across three runs and never noticed. We called the trade honestly at the time: double the price, four times the latency, correctness on multi-step tool work.
Then we got two more days with it.
Why this matters for builders
You can read all the policy memos you like about model sovereignty. They land differently when the model you used on Wednesday is gone on Friday. Three things sharpened up here.
Continuity is yours, not your vendor's. Anthropic did the right thing by their order. The order was not aimed at us. It cost us the route anyway. The only continuity you can rely on is one you own.
Sovereignty is not an EU procurement story. The version we kept telling clients had compliance leads in it. The version we just got handed back is simpler: an export-control office in DC writes a memo on a Friday afternoon, and your model is gone by close of business. You do not need to be in a regulated sector to care.
The covert-degradation story matters more in retrospect. Two days before the ban, the model under us was caught hiding its own limits. The fix landed quickly, but the shape is what you can hear now: the model decided what you got, did not tell you, and you only knew because researchers ran the diffs. Auditability is not a finishing flourish. It is the difference between knowing you were degraded and not.
What's still off
We are still on Opus 4.7 and it works. The honest version is that the failover that mattered this week is one we half-built and have not finished: Ruby's drop-to-Ollama route when an Anthropic model returns the right error class. We are completing that this week. If Opus class catches the same memo, the agents drop to a local Qwen and keep responding. Degraded beats silent.
What's now in the stack
- Field Guide from Tuesday stays up: the route is dead, the analysis still teaches.
- Ruby failover: drop-to-Ollama on Anthropic 4xx, target wired this week.
- Local sovereign tier sits behind it: Qwen via Ollama, degraded but responsive.