For nearly three weeks, the best coding model in the world did not exist. Not deprecated, not rate-limited. Off. On June 12 a US export-control directive forced Anthropic to switch Claude Fable 5 off for every customer on the planet, and it stayed dark until July 2.
If you had wired a workflow around it that first week, you did not have a slow model or an expensive one. You had nothing. That is the part worth sitting with, because the model is back now, and the temptation is to exhale and forget it happened.
The Model That Vanished
Fable 5 launched on June 9 as the strongest coding model anyone had shipped: first on GDPval, 95% on SWE-bench Verified, the clear frontier. Three days later it was gone. The Department of Commerce issued an export-control directive citing national security, and the way the rule was written, Anthropic could not simply fence off foreign access and keep the model live for everyone else. To comply, it had to disable Fable 5 and its unrestricted sibling Mythos 5 for all customers, everywhere.
The trigger was a jailbreak. Amazon researchers found a way around Fable 5's safeguards, prompting it to hunt for software vulnerabilities, and in one case it produced working exploit code. That finding is what set the export order in motion. Anthropic complied while saying plainly that it disagreed, and pointed out the uncomfortable part: the same demonstrations could be coaxed out of Opus 4.8, GPT-5.5, Kimi K2.7, Sonnet 4.6 and others. The capability was not unique to Fable. The directive landed on Fable anyway, and for the three weeks in between, the model was off.
Here is the part that should unsettle anyone building on a single model. On Anthropic's own severity scale, the reported Fable jailbreak was the mildest kind, a minor one that recovers benign work the classifier blocks by default, not a universal break that unlocks real harm. A minor, third-party finding still took the best coding model on earth offline for every customer on the planet. You do not get a vote.
It Comes Back Changed
On June 30 the Commerce Department lifted the controls, and Anthropic began redeploying Fable 5 globally over the next two days, broadly available again by July 2. But read the fine print on the return. The model comes back behind a new safety classifier tuned to block the reported technique in more than 99 percent of cases, and to buy that margin Anthropic widened the boundary so far that it now catches benign work too. In the near term, routine coding and debugging fall back to Opus 4.8 while the classifiers settle.
And now read the asterisk in the headline. Fable 5 was included in Pro, Max, Team and select Enterprise plans only through July 7, at up to half of weekly usage limits. Then the backlash landed, and on July 8 Anthropic moved the deadline once: five more days, through Sunday July 12 at 11:59pm PT, same 50% cap. From July 13 it comes off subscriptions entirely and lives outside usage limits altogether: it bills only through usage credits, at $10 per million tokens in and $50 out, the same meter the API has charged from day one. You switch credits on under Settings, load a balance (redemptions cap at $2,000 a day, auto-reload optional), and every Fable token draws it down. No credits enabled, no Fable. The API and consumption-based Enterprise plans are untouched.
Anthropic insists the removal is not permanent. Demand, it says, is "very high, and difficult to predict," and it aims to restore Fable as a standard part of subscriptions "as soon as capacity allows." No date attached. Days after the model came back, it leaves again for everyone who does not pay separately for it. If you are staying on Fable past the cutoff, the routing playbook for not paying full price is here.
So the thing that came back on July 2 is not quite the thing that left on June 12. It is Fable 5 with a heavier hand on the wheel, quietly routing some of your everyday work to a different, older model, and on a meter that starts July 13. Better than dark. Not the same as before.
The full model
Fable, with a governor
The Model Is a Tenant
Here is the lesson, and it has nothing to do with export policy. The most capable model you can buy is a dependency you do not control. It can be switched off by a directive you never saw coming, it can come back changed without asking you, and it can come back metered: eleven days after the return, it leaves subscriptions again for anyone who has not enabled usage credits. For three weeks, every team that had wired Fable 5 into the center of how they work was building on sand.
Some of them watched what filled the gap. While Fable was dark, an open-weights model, Z.ai's GLM-5.2, quietly took the top spot on Design Arena's web-design board, the first open model ever to do it. The frontier did not stop moving because the leader was unplugged. It just moved somewhere you could actually run.
Levie
"The big winner in all of this is going to be open weights models."
The teams that felt it most were the ones moving fastest. At Every, Dan Shipper had the crew racing the clock in the short window before the shutdown landed, treating a frontier model as a resource that might not survive the week.
Shipper
"BUILD EVERYTHING."
What This Means If You Are Building
The answer is not to bet on the other lab. Codex could be next. Any model behind an API can be. The answer is to stop treating the model as the thing you own. Own the layer around it instead: the context you feed it, the skills and workflows you have built, the map of your business it reasons over. That layer is portable. Point it at a different model on a Tuesday and keep working.
This is the whole idea behind an operator layer, and it is what we install on site. Build so the model can disappear, because sometimes it does. Keep an open-weights option warm so you are never one directive away from dark. Rent whatever model is best this month. Own the loop that makes it useful.
Fable 5 is back, and it is still, on its day, the best in the world. Use it. Just do not build your company on the assumption that it will be there tomorrow.
How It Actually Ended
On July 20 the countdown stopped. Anthropic made Fable 5 a standard part of Max and Team Premium, included at up to 50% of weekly limits, with no expiry attached. Pro and Team Standard went the other way and now reach it through usage credits, cushioned by a one-time $100 credit.
The reason was supply rather than a change of heart. The SpaceX Colossus 1 deal added more than 300MW and over 220,000 NVIDIA GPUs to what Anthropic could draw on, and efficiency work brought down the per-call cost of frontier inference. The constraint that produced three moving deadlines got relieved, so the deadlines stopped moving.
Four days later Anthropic shipped Opus 5 at $5 in and $25 out, close to Fable on most work at half the price, and unlike Fable it is compatible with zero data retention. For a lot of teams that had spent June and July rearranging their stack around Fable access, the practical answer arrived from a different direction entirely.
Which is the point the whole episode keeps making. Three deadlines, one reprieve, one permanent inclusion, one tier demoted to credits, and a cheaper model that resolved the problem sideways, all inside seven weeks. None of that was predictable in June, and none of it is the kind of thing you can plan a company around.
Rent the model. Own the loop.
Own the layer, not the model.
Book a free Diagnostic: 30 to 45 minutes, no deck, no pitch. We map what you have built on top of a single model, and where an operator layer you own, the context, skills, and workflows, would keep you running when the model does not.
Book the Diagnostic →