Anthropic shipped Claude Fable 5 on June 9. Three days later the US government applied export controls, and Anthropic suspended access for every user because it had no reliable way to verify nationality in real time. Access came back on July 1, after the controls were lifted on June 30.
Nineteen days. If a workflow you depend on was pinned to that model, you had a nineteen-day outage you did not cause and could not fix.
Anthropic handled it about as well as a lab can: a public explanation, restored access, usage credits for paid plans. The lesson lands anyway. A frontier model is now an uptime dependency, the same category as your DNS provider or your payment processor. Almost nobody manages it that way.
Three ways a model disappears
The hard outage is the dramatic version. There are two quieter versions that hit more businesses more often.
| Change | Recent example | What breaks |
|---|---|---|
| Hard outage | Fable 5 offline June 12 to July 1, 2026 | Every workflow pinned to it, immediately |
| Deprecation | Labs retire older models on schedules measured in months | Workflows nobody remembered to migrate |
| Economics shift | GPT-5.5 input pricing doubles past roughly 270K tokens of context | Budgets built on last quarter's rates |
There is also a fourth, alias drift. If your code calls a rolling model alias instead of a dated snapshot, the provider can change what answers your requests without telling you. The behavior shifts under prompts that were tuned for the old behavior, and the first sign is usually a confused customer.
What this looks like on our own site
The concierge on this site runs on OpenAI. The model is named in exactly one place, an environment variable with a safe default:
// src/lib/agents/config.ts (abridged)
model: process.env.OPENAI_CONCIERGE_MODEL ?? "gpt-5.4",
rateLimitMax: intEnv("AGENT_RATE_LIMIT_MAX", 8), // requests per minute per IP
dailyTokenBudget: intEnv("AGENT_DAILY_TOKEN_BUDGET", 50_000),In production that variable currently pins gpt-5.4, so the default in code is exactly what it should be: a fallback that never decides anything. Every run writes a receipt into our CRM with the model ID, prompt hash, and token counts. If that model went dark tomorrow, the swap is one config change and a redeploy, and the receipts show exactly which conversations ran on which model before and after the switch.
Honest constraint: that surface has one provider today. A visitor concierge can tolerate an hour of downtime while I change a variable, so paying to keep a second provider warm is not worth it there. A workflow that answers customer calls or writes to business records has a different tolerance, and it deserves a tested second route before it needs one.
The dependency audit
If question five comes up empty, start there. You cannot manage a migration you cannot see, and that visibility is most of why receipts exist.
How to pick the fallback
Skip the leaderboards. Take twenty real cases from the workflow, run them through the candidate model with the same tools and the same rubric, and count what changes: retries, human corrections, cost per accepted output.
Mid-2026 makes this cheaper than it used to be. GPT-5.6, Gemini 3.5 Flash, and Fable 5 all handle tool calling well enough that the harness code barely changes between them. The differences that matter will show up in your twenty cases, and they will be specific to your work.
Price the fallback while you are at it. Current list rates run from under a dollar per million output tokens on small models to $50 per million on Fable 5. A fallback that costs five times the primary is fine for a one-week outage and a problem as a permanent home.
Treat it like vendor management
Model choice used to feel like a technology decision. It is a vendor decision now, and this June proved the vendor's regulator is part of the release cycle. Labs will keep shipping fast, repricing, retiring old versions, and occasionally going dark for reasons nobody predicted.
None of that is in your control. The audit is. One page, six questions, reviewed quarterly, next to the insurance renewals where it belongs.
