Fable 5, Rationed.

Originally published in Temrel, a weekly newsletter on agentic engineering.

Imagine someone tosses you the keys to a Ferrari, then mentions the tank is half full and the pumps close Monday. That is Fable 5 this week. The most capable Claude yet just came back from the dead, and it showed up rationed.

Which makes the interesting question not "is it good." It is "what do you actually spend it on."

The most capable model Anthropic has shipped just came back, capped

On June 30, 2026, the US Commerce Department lifted its export controls on Claude Fable 5 and Mythos 5. Fable 5 returned globally on July 1, across the Claude Platform, Claude.ai, Claude Code, and Cowork (CNBC, 9to5Mac).

There is a catch, and it is the whole point. This is a limited release. Pro, Max, Team, and select Enterprise get Fable 5 at 50% of normal usage limits through July 7. After that it moves to usage-based credits. Translation: the best model you have access to is capped today and metered next week.

Anthropic pitches Fable 5 as more capable than Opus 4.8. Their numbers: FrontierCode Diamond at 29.3%, against Opus 4.8 at 13.4% and GPT-5.5 at 5.7%, plus a 50-million-line Ruby migration done in a day (Anthropic, Vellum). Say the quiet part first: those are Anthropic's own figures, vendor-reported, not independent benchmarks. Treat them as a marketing floor, not a measured ceiling. Then notice that even if you halve them, the gap is still real.

It was gone for a reason, and that reason is the ending

Quick backstory, because it comes back around. Fable 5 was pulled on June 12 after an Amazon report showed a prompt could bypass its safeguards and get it to surface software vulnerabilities. Anthropic says a new classifier now blocks that technique in more than 99% of cases. Hold that thought.

The skill this week is triage, not hype

Here is the thing nobody selling you a model will say out loud: a capped, soon-to-be-metered model turns capability into a budgeting problem. When the good stuff is finite, the skill is deciding what deserves it. Not "how do I use the new model." "What do I aim it at."

So, a rule. Spend Fable-5-grade capability on work that is long-horizon, high-ambiguity, and codebase-wide:

• The migration nobody wants to start.
• The refactor that touches forty files and needs a plan before a single line moves.
• The bug that has already outlived three engineers.

That is where a genuinely stronger model earns its credits, because that is the work where the difference between "good" and "best" actually changes the outcome.

Everything else stays on Sonnet 5. The endpoint. The unit test. The rename. The "write me the boilerplate." Routine, well-specified, low-blast-radius work does not get better with a Ferrari. It just gets more expensive.

The reason this matters more than the usual model-launch noise: the delegation gap. Anthropic's own 2026 Agentic Coding Trends Report found developers now use AI on roughly 60% of their work but fully delegate only 0 to 20% of tasks (report). We reach for the model constantly and hand it the whole job almost never. Fable 5 is pitched as the thing that shrinks that gap, the model capable enough to take an entire task and hand back something you would actually merge. If that is even half true, it is exactly the work you want it on. Do not spend your rationed capability autocompleting functions you could have written in your sleep.

The thing that makes it dangerous is the thing that makes it useful

Now the part that should sit with you. The reason Fable 5 got export-controlled is the same reason it is worth rationing: it is capable enough to autonomously find software vulnerabilities. The property that made regulators nervous and the property that makes it do your hardest refactor unsupervised are not two things. They are one thing.

So capability and safety are not opposed here, they are coupled. A model strong enough to take the whole task is, by construction, strong enough to take a task you did not give it. That classifier now catching the exploit in 99-plus percent of cases is not a footnote. It is the cover charge for a model this strong being allowed out the door at all.

Spend it like it is what it is: powerful, finite, and not entirely tame. None of that is a reason to avoid it. It is a reason to aim it.

Subscribe to Temrel for weekly agentic engineering field notes.