Fable 5, Opus, or Sonnet: which Claude for client work

Update — July 3, 2026. One more change to the picker: Anthropic has announced that Fable 5 stays on paid plans only until July 7, 2026 — capped at up to 50% of a plan's weekly usage until then — and after that is available only through the developer API, billed per token. A return to subscription plans has been left open but not committed to. After July 7, the in-app picker this article is about reads Opus and Sonnet for everyone, and the Fable 5 rows in the table below route to Opus — exactly as the leash rule says they should.

Update — June 26, 2026. After this published, Anthropic restricted Claude Fable 5 access following new US export controls; it's currently unavailable to non-US nationals, and the rules are still moving. If Fable 5 isn't in your picker, the choice below is simply Opus vs. Sonnet — which is how everyone chose two weeks ago, and the work shipped fine. The leash rule that drives this whole piece doesn't care what the top model is called: route long-leash, whole-deliverable work to the most capable model you can reach (Fable 5 if you have it, Opus if not) and route the short-leash work down. On building a practice that survives losing a model: Don't build your practice on one AI model.

Fable 5 reshuffled the top of the Claude model picker, and that makes the picker a real decision again. Until recently the move was simple: choose the most capable model, leave it there, forget the menu exists. Now the most capable model on your list is also the slowest to answer and the deepest thinker on it — and for half of what a consultant does in a day, that depth is wasted on the task.

Most of the content answering "Fable 5 vs Opus" is written for developers: benchmark charts, API prices, context-window trivia. None of it tells you which model should write Friday's client update. So here's the consultant version. One rule, then defaults, task by task.

What's actually in the picker

Skip the spec sheets. Here's the working translation.

Fable 5 is Anthropic's new top model — a tier above the Opus line, billed as their most capable model (access now restricted to US nationals, and on paid plans only until July 7, 2026 — see the updates above; if it isn't in your picker, your top tier is Opus, described next). Its differentiating strength is long, multi-step work: it holds an entire deliverable coherent from start to finish instead of drifting in the middle, and it reasons before it answers, every time. The trade is time — responses to hard asks take noticeably longer, because the model is doing more before it speaks. We covered what it changes for consulting work on Friday; this piece is about when to choose it.

Opus was the flagship until last week, and it's still an excellent model — the one most serious Claude users have been running for months. Strong judgment, strong writing, fast enough for genuinely interactive work. Nothing about Fable 5's arrival made Opus worse at anything it was doing for you on Monday.

Sonnet is the speed-and-intelligence balance: noticeably snappier in conversation, and more than smart enough for work whose shape you already know.

If your picker shows a lighter option such as Haiku, file it under instant answers — lookups, quick transformations, "make this a bullet list." Most consultants will rarely choose it deliberately, and that's fine.

Two practical notes before the rule. First, cost: on a paid plan, these live inside the subscription you already pay — Fable 5 only until July 7, 2026, and capped at half your plan's weekly usage until then (see the updates above). Plans meter usage, and the bigger models draw down that allowance faster. Spending Fable 5 on a subject line is paying caviar prices for toast. Second, names: Anthropic ships new versions on its own schedule, and the labels will keep changing. The tiers persist — a top model for the hardest work, a strong default, a fast one. Learn the tiers and the rule below survives every release.

The rule: match the model to the leash

Forget capability rankings. The question that actually picks the model is this: how long is the leash on this task, and what does wrong cost you?

Leash length is how much work you hand over between check-ins. A subject-line rewrite is a ten-second leash — you see the output immediately, a miss costs nothing, you iterate without thinking about it. A "here's the transcript, my services file, and my rate floor — draft the proposal" run is a long leash: the model works through many steps before you see anything, and a failure in the middle compounds quietly until the end.

Put those two on the same model and you've made an error in one direction or the other.

The second half of the rule — what wrong costs — decides the borderline cases. Cheap wrongness gets caught for free: you were reading the output anyway, and the fix is one more message. Expensive wrongness is the kind that survives your review because the artifact looks finished — a proposal with a plausible-but-off price, a case study citing a detail the client never said. The polish of a long-run output is camouflage. The more finished a draft looks, the stronger the model you should have used to make it, and the harder your read should be before it leaves your desk.

Short leash, low stakes → the fast model. You're in a tight loop. Latency is the dominant cost, and you personally inspect every output anyway. Depth has nowhere to earn its keep.

Medium leash, judgment-gated → the strong model you steer. The work happens in stages and your judgment gates each one — scope before price, price before draft. You want maximum sharpness per stage and a loop fast enough not to break your momentum between them.

Long leash, whole deliverable → the model built for it. This is where Fable 5 earns its slot. Delegation runs reward the model that can hold the entire job in its head — and the middle of a long run is exactly where weaker models used to drift.

Concrete contrast: the Friday client update is a short-leash task wearing business clothes — known shape, your inputs, your proofread before it sends. The proposal built from a discovery transcript is a long-leash task wearing the same clothes — a dozen quiet decisions about scope, sequence, and emphasis happen between your brief and the draft. Same client, same week, opposite ends of the picker.

That's the whole rule. Everything below is the rule applied.

Task by task

Default assignments for the recurring work of a solo practice. Adjust to your taste — the point is to have defaults instead of re-deciding at every blank conversation.

Task	Default	Why
Email passes, rewrites, quick reformats	Sonnet	Ten-second leash. Speed wins.
Friday client update from the week's notes	Sonnet	Known shape; you review before sending anyway.
Discovery call debrief	Opus	Steered extraction with judgment. Go Fable 5 when the transcript is long and messy.
Scope and pricing options you steer step by step	Opus	Your judgment gates each stage; sharp steps, fast loop.
Proposal draft from a full brief	Fable 5	The whole-deliverable handoff — the longest leash you own.
Case-study draft from a project folder	Fable 5	Assembly across many sources; coherence over length.
Long-document work (a 40-page RFP, a contract you must absorb)	Fable 5	Sustained attention across a long input is the headline capability.
Thinking partner, "what am I missing here"	Opus	Fast enough to argue with. Depth without the wait.

Three patterns worth noticing. Everything routine routes to Sonnet. Everything judgment-steered routes to Opus. Everything delegated-whole routes to Fable 5. When a task doesn't fit a row, ask the leash question and it sorts itself. (If Fable 5 isn't available to you — and after July 7 that's everyone working in the app; see the notes up top — the delegated-whole work routes to Opus, your top tier instead; the rule holds, only the name in those rows changes.)

The Friday-update default surprises people — surely the better model writes the better update? Marginally, maybe. But a good weekly update comes from its shape and its inputs, not from model depth, and you proofread it regardless. Save the depth for where wrongness is expensive.

The proposal row is the one that actually moved this month. Pre-Fable, the proposal was a steered task — you chained the stages because no model could be trusted with the whole thing at once. With a model built for long runs, it becomes a delegated task for consultants who can write a complete brief: objective, inputs, constraints, what done looks like. If your handoffs aren't that complete yet, keep steering on Opus — a long leash on a vague brief just produces a longer wrong document.

And the discovery debrief deserves its own note, because it's the highest-stakes steered task on the list: what you extract from that call feeds the scope, the price, and the proposal that follows. If you want a tested starting point rather than a blank page, our free Discovery Debrief is a single-prompt workflow with a Claude skill and a worked example — it runs the same whichever model you pick, and it costs nothing.

When Fable 5 is the wrong choice

The new model deserves a section of honest no.

When you need the loop, not the leash. Interactive drafting — reacting line by line, talking a document into shape — wants the fastest model that's good enough. Fable 5's deeper reasoning shows up as wait time, and waiting breaks the rhythm of steered work. A brilliant answer ninety seconds late is a worse partner than a very good answer now.

When the shape is known. A weekly update, a meeting recap, an invoice reminder: structure plus your inputs solves the task. Extra capability has nowhere to go, so you're paying its costs — time, usage — for nothing.

When you're rationing usage. Heavy Fable 5 runs draw down plan allowances faster than the smaller models do — and for as long as it stays on plans, Anthropic caps Fable 5 at half your weekly usage outright. If month-end finds you bumping into limits, audit whether routine tasks crept up-model out of habit. They usually have.

When you can't get it. Access has narrowed twice: US export controls made Fable 5 unavailable to non-US nationals, and Anthropic has now put an end date on in-app access — after July 7, 2026, it leaves subscription plans entirely and is available only through the developer API, billed per token. A return to plans has been left open, with no specifics. So after July 7, everyone working in the app defaults to Opus and Sonnet — which were everyone's defaults a month ago, and the work shipped fine. The leash rule is unchanged: your long-leash, whole-deliverable work routes to Opus instead, and the table above reads with Opus in the Fable 5 rows.

One footnote so it doesn't surprise you: Fable 5 ships with safety classifiers around a handful of research-heavy topics — security research, biology — and Anthropic says fewer than five percent of sessions ever encounter them. Consulting work essentially never goes near the triggers. If a conversation ever behaves oddly, Friday's piece covers the practical response: fresh conversation, plainer ask, move on.

Set your defaults this week

Three moves now, one for the calendar. None longer than a coffee.

Write the defaults down. List your five most frequent Claude tasks. Assign each a model from the table above. Put the list somewhere you'll see it. The win isn't the assignments — it's never re-deciding at a blank conversation again.

Run one honest A/B. Take a deliverable you produce regularly. Same inputs, same brief, two models — your current habit and the table's recommendation. Read both drafts cold and pick the one you'd actually send. That's your evidence, from your practice, not a stranger's benchmark screenshot.

Re-route the routine work down. If everything has been running on the biggest available model out of habit, move the known-shape tasks to Sonnet for a week. You'll notice the speed immediately, you almost certainly won't notice a quality drop, and the usage headroom comes back for the long runs that genuinely need it.

And the calendar one: put an expiry on your defaults. Models change under you, and assignments go stale quietly — the table above would have read differently a month ago and may read differently by fall. When the next release lands, re-run the A/B on your highest-stakes task before believing anyone's writeup, including this one. Ten minutes, evidence refreshed.

The deeper question underneath all of this — which work to hand over at all, and which work stays yours no matter how capable the models get — doesn't change with the picker. That's the operating frame of The Solo Consultant's AI Playbook, and it's deliberately model-agnostic. Pick the model per task. Keep the judgment on retainer.