Claude Fable 5 is back — but on a leash. Five days, half your weekly limit, and it now refuses routine coding and security work. Here's how to use it before July 7, and the setup that keeps you in control after.
Fable 5 returned to the Claude model picker — but they had to cripple the best model just to be allowed to give it back. Here's the fine print.
You only get Fable 5 for up to half your weekly usage limit. Hit 50% and you drop back to Opus or Sonnet.
After the 7th, Fable 5 leaves your subscription and moves to pay-as-you-go usage credits — reportedly ~$10/M in and $50/M out, roughly double Opus 4.8.
A widened safety classifier now blocks more benign requests — routine coding, cybersecurity, biology — and quietly reroutes them to Opus 4.8.
Fable was pulled offline mid-June and only returned in July. A model you depend on can be turned off — and back on with new rules. You don't get a vote.
Fable is overkill for most tasks. Use a cheaper model for setup, keep the effort setting low, babysit it — and reserve it for deep, high-value work.
Point it at your projects and memory: "You're the most capable model I have — tell me the handful of tasks genuinely worth running by you." That's your shortlist.
Give it a plan doc plus your real data through connected tools, and ask for an honest 3-month focus. Advice that changes your whole quarter.
Hand it the whole codebase: "Find everything wrong before I ship." In one demo it spun up agents and found 12+ real bugs — including a user-data leak.
Fable writes the detailed plan and architecture — clear enough that a cheaper model executes it step by step. Expensive model thinks, cheap model types.
Stripe reportedly migrated 50M lines in a day with a Fable-class model. Point it at your messy repo or your own drifting systems and let it clean house.
The best AI isn't getting more open — it's getting more gated. GPT-5.6 needs approval to access. Fable comes back on a leash. And everything here is going to get regulated, which has an effect. The smart move isn't chasing this week's frontier model. It's owning your setup. In my opinion, the next best option — the one nobody can revoke — is running local, and the alternatives ladder goes from free to paid.
Runs on your own machine. Private, unlimited, can't be switched off. Start with the laptop in your closet or a Mac mini. Grab Ollama or LM Studio and pull Qwen3, gpt-oss (yes, OpenAI open-sourced one), or Gemma. A small model runs in ~16GB.
NVIDIA gives away a strong open model — Nemotron 3 Ultra — through a free developer API. No subscription, coding-competitive with GPT-5.5, open weights. A serious option when you want near-frontier quality without paying anyone. (Verify the free tier before you rely on it.)
A frontier-competitive model out of Japan you drive from your own machine via a local CLI, paying per use. Not sovereign like fully local — you're still calling an API — but it's another door that doesn't depend on the same handful of US players.
No per-token toll and no rate limits means a whole class of work becomes possible — the always-on stuff you'd never run on a metered cloud model.
Use Fable 5 while it's cheap to design your systems — the plans, the architecture, the agent blueprints. Then run them on hardware you own, with models nobody can revoke. Frontier for thinking. Local for doing. That's how you grab the advantage now and actually keep it.
To orchestrate all of it you'll want an agent layer — Hermes agent or OpenClaw. A straight comparison of the two is coming next.
The exact local setup, the models I run, the agent templates, and a community working through all of this in real time — it's inside the AI Founders Vault.
Join the AI Founders Vault →