On Tuesday, Anthropic released the most capable AI model ever made public. By Friday, the US government had ordered it switched off — worldwide, for every customer. That’s the whole week in one story: you don’t own your AI supply chain — you rent access to it, and the lease can be torn up by people who never signed it. Here’s what happened, and the three things to do about it before it happens to a tool you actually depend on.
The most capable model anyone had shipped
On June 9, Anthropic launched Claude Fable 5, its first publicly available Mythos-class model and, by its own account, the most capable it has ever shipped. The benchmarks backed the claim. Fable 5 posted 80.3% on SWE-Bench Pro — Anthropic’s hardest agentic software-engineering test — roughly 11 points clear of the next-best frontier model and well ahead of Opus 4.8 at 69.2%, OpenAI’s GPT-5.5 at 58.6%, and Gemini 3.1 Pro at 54.2%. It hit 95% on SWE-bench Verified and 88% on Terminal-Bench, the evals that measure whether a model can carry a long, multi-step task to completion on its own.
But the number that tells you what it actually felt like is this one: in a published run, Fable 5 carried out a codebase-wide migration across 50 million lines of code in a single day — work a full team would have spent more than two months doing by hand. What changed wasn’t raw knowledge; it was persistence. The model kept going, trying alternative approaches instead of stopping to hand the problem back — the autonomous, long-horizon execution that does real work rather than just answering questions.
It shipped at a premium price — $10 per million input tokens and $50 per million output. On paper that is roughly double the per-token cost of the model it replaces, but Fable 5 is faster and noticeably more sparing with tokens, reaching a good answer in fewer of them. Run a real task end to end and the cost comes out more like a third to a half higher, not double — a modest premium for a marked jump in quality. It was free to plan subscribers through June 22.
Here today, gone Friday
Three days later, at 5:21pm Eastern on June 12, a US government export-control directive landed. Citing national security and an alleged jailbreak, the order barred access by any foreign national anywhere — including Anthropic’s own staff — so the only way to comply was to disable Fable 5 and Mythos 5 for everyone, everywhere. Anthropic disputes the rationale, says the technique only surfaced minor flaws that other public models can find too, calls it a misunderstanding, and is working to restore access. In the meantime it is issuing refunds.
Sit with the shape of that. The single most capable model on the planet on the day it shipped — gone three days later, not to a competitor, not to a bug, but to a federal order the vendor didn’t see coming and doesn’t agree with.
Here’s the part that should change how you build. If your firm had wired a client deliverable to Fable 5 during those three days, it vanished on Friday with no deprecation notice, no migration window, and no appeal. That’s the institutional-constraint problem in its sharpest form yet — not vendor lock-in, not a safety toggle you can argue about, but a sovereign actor removing a tool overnight. And notice what did not break: every other Claude model stayed live, and Anthropic’s own design quietly fell back to Opus 4.8 for the tasks Fable blocked. The vendor built a fallback into its own product; the question this week is whether you’ve built one into yours.
And that was the loud version
While Fable was making headlines, the same thing was happening quietly across the rest of the stack — on the vendors’ schedules, not yours.
OpenAI churned its lineup in the same days: GPT-5.2 was pulled from ChatGPT on June 12, with live conversations auto-migrated to GPT-5.5, GPT-4.5 retires on June 27, and o3 follows in August. Google enabled Gemini 3.5 Flash by default and made it non-disableable for Gemini Enterprise users from June 8 — if your model exposure comes bundled with your productivity suite, you didn’t choose the model, the platform did. And it isn’t only the model layer. Intuit retires classic QuickBooks reports on June 15 — modern view becomes the only option — which is the same mechanic landing in the general ledger your team touches every day.
Line them up and the pattern is an escalation: a sovereign actor pulled a model, vendors are deprecating models on their own cadence, and your core platform is sunsetting a feature you may still rely on. Fable was the dramatic version. This is the everyday version, and it’s relentless — you will be migrated whether your quality control has caught up or not. The lesson is the one I’ve made before: a single-vendor, single-version strategy isn’t a safe harbor, it’s an unhedged position.
Quick hits
OpenAI is buying Ona to run agents while you sleep. OpenAI announced it will acquire Ona — formerly Gitpod — to give Codex secure cloud execution so agents can run for hours unattended, “while your laptop is closed.” Strip the developer framing and the direction matters to you: agentic work is moving from something you babysit in a chat window to persistent, server-side execution. The more work runs without you watching, the more your verification layer earns its keep.
Scaling New Heights opens in Orlando. The Woodard conference runs June 14–17, with Microsoft joining as a first-time sponsor — a Monday keynote and a full AI training track, a clear signal of where the platform giants think the profession is heading. I’ll be on the floor all week. If you’re there, message me on LinkedIn — I’d like to meet.
What this week actually decided
Step back and the week resolves into a single procurement question you can apply to any AI tool you adopt from here: if this vanished on Friday, what would break, and what’s my fallback? Continuity has quietly become a buying criterion, ranked alongside capability and price — because capability you can lose overnight isn’t capability you can build a practice on.
Before you over-correct, notice what did not break this week. Your bookkeeping ran, your close ran, your advisory prep ran. None of it lived on the bleeding edge, and the models underneath it stayed up. This is a call for hygiene, not panic — and the hygiene is cheap.
I wrote the playbook for exactly this back on April 29: three tests any firm can stand up in an afternoon — drift detection, portability, and a version trail. You can read that piece on our blog at theaiaccountant.ai, and download the companion Vendor Test Pack from the page — it hands you the exact prompts to run all three. The test that matters most this week is portability: can you run the same workflow on a different model tomorrow? If the answer is no, that’s the work — before the next directive, deprecation, or sunset decides it for you.
That decision — what your practice should depend on, and what it should be able to swap — comes before any tool choice, and it is exactly the work our AI Practice Transformation program is built for. The next cohort starts June 26: four weeks, one three-hour session a week, working through how your firm operates, prices, and competes in this AI world. Details and a call with me at theaiaccountant.ai/transformation.