Fable: here in an instant, then gone
Last week I wrote about Fable 5 — Anthropic’s newest model — and the catch: it was only on the normal subscription plans for a short window. I said I’d build something with it that night. And I did, and just in time, too. Because it’s already gone.
On Friday the 12th, the US Commerce Department handed Anthropic an export-control directive citing national security — ordering it to cut off Fable 5 and its ungated sibling, Mythos 5, for any foreign national, anywhere in the world. There’s no way to tell who’s a foreign national in real time, so Anthropic disabled both models for everyone, the same day. The model Anthropic called its most capable yet was live for a matter of days.
The US government’s stated worry is a “jailbreak” — a way to coax cyber-vulnerability help past the model’s safeguards. Anthropic says the holes that demo turned up were minor and already known, that other models find them too, and that the whole thing is a misunderstanding that it’s working to reverse.
But for the short time that it was available, I gave it one job: build a visualisation of my home network as a solar system - the Network Orrery
I let Fable create the plan first, which invoived running the whole build itself across a set of smaller agents, and kept the real network data — the MAC addresses, the device names — out of its own sight the whole time, on purpose. What came out was a deployed site that turned a day of DNS traffic into constellations and supernovae, with nothing leaked.
One thing worth stealing from how it got built: when it came to the look, I didn’t describe it. I handed it five reference videos from Dilum Sanjaya for inspiration and one line — make it feel like one of these. It pulled frames out of the videos with ffmpeg (a local, free tool), studied them, and came back with the aesthetics on the table; I picked one and it re-themed the whole thing to match.
Then it journaled the build in first person, with the bits it got wrong left in. Mid-build, it told me with complete confidence that it was past midnight and into the next day. It was 23:38. It only corrected itself when I told it to go check the clock — the exact rule its own instructions carry. The model that tops every benchmark still couldn’t read a clock, because it guessed instead of checking. That’s the whole case for keeping a human in the loop, and for the discipline mattering more than the horsepower.
Near the end, it recommended replacing itself with a fresh copy that wouldn’t remember any of this, and wrote that copy a set of instructions detailed enough to finish the job blind. The new one picked it up and shipped it.
I’ve put the whole journal up on the orrery site — unedited, mistakes left in (I changed a couple of details to keep family and network specifics private; nothing else). It’s a fair record of what the model could do in the few days it was available, written in its own voice.
What it cost — and who actually paid
Separate from the government order there’s one more thing worth knowing. When I closed the build out, I had Claude tally the compute from the session logs and price it at the pay-as-you-go API rate. The Orrery — three days, all those agents — came to about $423. My whole week, across every project, was around $4,200 of compute. I don’t pay that. I pay a flat monthly subscription — the same one you can buy. The gap is the subsidy: the real cost of this stuff is being absorbed so it reaches you cheap. (Fable made it vivid — it counted double against your plan limits.)
That’s not why it got pulled — the US government was. But it’s the same lesson from the other side: the terms underneath these tools — who’s allowed to use them, and what they really cost — can change with no notice.
The model’s gone (for now), but the way of working that made the build hold up — the planning, the review, the handoff — is still here. That’s the part worth keeping in mind no matter what models you work with.
— Jim
€7.00 / month
Signal Over Noise
Skip the AI Hype.
A newsletter about using AI to actually run a business and build things — past the hype and the doom,… Read more