Working with Fable: you have until July 12.

Anthropic just extended included access to Claude Fable 5 on paid plans through July 12. After that it moves to usage credits and returns to subscriptions “as capacity allows,” which is vendor language for “don’t count on a date.” So there’s a dated window, and a question worth answering before it closes: what are the best use cases I can throw at it now?

The easy mistake is spending the window using Fable to do Opus’s work. You can do that, of course, and in many cases you’ll get better output. But the real value is in finding the work that Fable is exceptionally good at, and other models fail at completely.

In The AI Usage Ladder I argued that the level is a property of the problem, not the person. The same rule applies to model tiers. Fable is a new top rung, and the question is which of your problems need it, not whether it’s better (it is). Anthropic says the same thing in its own prompting guide: teams get the best outcomes applying Fable to their hardest unsolved problems, and testing it on simpler workloads undersells it.

01 · The modelWhat Fable is, in plain language.

Fable 5 is Anthropic’s new top model tier, a class above Opus. It launched June 9, disappeared June 12 under a US export-control order, and came back July 1 with tightened safeguards. Through July 12 it’s included in Pro, Max, Team, and select Enterprise plans, for up to half of your weekly usage limit.

What the tier buys you, per Anthropic’s own documentation: it stays productive across multi-hour autonomous runs, it gets complex, well-specified builds right on the first pass more often, it reads dense screenshots and documents more accurately, it finds more real bugs in code review, and it manages parallel subagents dependably instead of losing the thread.

The honest caveat: Fable ships with the heaviest safety classifiers Anthropic has ever put on a model, aimed at offensive cybersecurity and biology. Anthropic set the trigger threshold wide on purpose, so benign security-adjacent work (vulnerability triage, debugging around auth code) can get blocked, and the request silently falls back to Opus 4.8. If you do security work, expect false positives and budget patience for them. This is the tradeoff that let them ship the model broadly at all, and after watching the export-control episode play out, I understand the choice. It’s still friction, and you should know it’s there before you route a security task through it.

02 · The problemGive it a problem you’d normally not give a model.

Anthropic’s phrase for this is “start at the top of your difficulty range.” Pick something that takes a person days, not minutes: a migration review, a cross-repo bug that’s survived two previous attempts, an end-to-end module build against a real spec.

A couple things we’re using it to tackle:

• The end-to-end build of a new module, from research to specifications, implementation, testing, and delivery. Fable acts as the orchestrator, doing the complex tasks and offloading simpler ones (coding to the spec, for example) to Opus.
• A full refactor of an entire codebase for one of our internal tools.

03 · The promptPrompt for outcomes, then get out of the way.

State the goal and the success criteria. Don’t script the steps. Fable plans better than your step list, and a step list written for a weaker model actively constrains it.

Give the reason along with the request. “I’m working on X for Y; they need Z” changes what the model attends to, because context lets it connect the task to what matters instead of guessing at intent.

Then expect longer turns. Single requests can run many minutes. Autonomous runs go for hours. This is the biggest adjustment in practice: you stop watching it type and start checking in on it the way you’d check in on a colleague, asynchronously, against evidence.

Two instructions from Anthropic’s prompting guide are worth stealing verbatim. First, tell it to audit progress claims against actual tool results before reporting; in Anthropic’s testing this nearly eliminated fabricated status reports. Second, state boundaries explicitly: when you ask for an assessment, say the deliverable is the assessment, so it reports findings and stops rather than fixing things you only asked it to look at.

04 · The subtractionDelete instructions, don’t add them.

This is the counterintuitive one. Most of us have accumulated prompts, skills, and CLAUDE.md files tuned for older models: guardrails against failure modes those models had, step-by-step recipes for tasks they couldn’t plan themselves. Instructions written for old models keep Fable behaving like an old model, and overly prescriptive scaffolding degrades its output.

The window is a cheap way to audit your own process debt. Run a task with your usual scaffolding, then run it again with most of the scaffolding removed, and compare. Every instruction the model no longer needs is one you were paying for in tokens, in maintenance, and in capability ceiling.

05 · The meterMind the meter.

The 50% weekly-usage cap shapes everything above. Fable at high effort burns quota fast: high is the sensible default, xhigh is for the work where being wrong is expensive, and medium performs better than you’d expect. We’ve found lower effort on Fable often beats the highest settings on prior models.

So route the routine work down-ladder. Our own practice is to pin mechanical work (formatting, extraction, rote transforms) to Opus or even Sonnet subagents and reserve the expensive tier for planning, review, and decisions that are costly to get wrong. That’s the Usage Ladder’s “lowest rung that reliably solves the problem,” except now a hard budget makes the discipline non-optional.

Our own telemetry from Fable’s first week back (July 1–7) shows what the meter does. Adoption was fast: Fable went from 3% of our daily spend on redeploy day to 62% by the first weekend. For the week it took 11% of our tokens and 26% of our estimated spend at API rates, roughly $2,700 of $10,500. Sticker price is double Opus ($10 in / $50 out per million tokens, against $5 / $25), and the disproportion follows.

The session view is where it gets interesting, and it’s the chart at the top of this piece. We ran 41 Fable sessions with a 45-minute median, against 366 GPT-5.5 sessions with a 9-minute median. Fewer, longer, harder sessions is what the top rung looks like when the routing works. And per session the premium is smaller than the sticker suggests: a Fable session averaged $39 to Opus’s $33, because the sessions were scoped to work that deserved the tier.

Sonnet 5 nearly didn’t make the chart, and the reason turned out to be the best data point of the week. It shows up in only eight interactive sessions, three minutes median, about $4 each. Where it actually lives is inside four always-on orchestrator sessions that ran five to seven days straight: Sonnet 5 handles the high-volume events while Fable and Opus take the expensive turns inside the same session. One of those sessions logged 8,670 events for $1,797, with the cost split almost evenly three ways between Sonnet, Fable, and Opus. Nobody planned that as a case study; it’s just what the ladder looks like when someone builds a long-running agent and pins each rung to the work it’s suited for.

The same telemetry also called us out. We preach routing mechanical work down-ladder, and our meter says Haiku barely registered all week (18 sessions, $31 total) while the grunt volume ran on Opus and GPT-5.5. The lowest reliable rung is a discipline, not a default, and we’re still learning it too.

06 · The ruleThe decision rule.

When the model comes back “as capacity allows,” you’ll know exactly what it’s for, and what it isn’t, and you won’t be spending usage credits to rediscover it.

One more thing the past month taught, and I’ll keep it small: this model launched, vanished under an export order, and came back metered, all within four weeks. Whatever your AI process is, build it so the model tier is a swappable part.