00 Start

Claude Fable 5.1 is out. The headline isn't "it's smarter"

Anthropic shipped a new model on September 1. I dug through what actually matters — no benchmark porn.

Every release gets sold to us as "X% smarter." Boring and useless. What's interesting about Fable 5.1 is different: it works longer on its own (without you babysitting it every five minutes), it's cheaper on real tasks at the same list price — and it has one main dial that decides more than any prompt tweak.

This guide is not a press-release retelling. We'll test all three claims on your own task.

What you'll walk away with

• A marathon task handed to the model end-to-end — with a prompt built for it
• The effort dial: when to turn it down and save, when to turn it up
• Two writing fixes: the new model writes better but denser — cured with one line

00 Start

The three numbers worth knowing (the rest is noise)

Out of all the release benchmarks, these are the ones that matter for you and me.

Agentic research (Terminal-Bench-Science): 24.7% → 52.6%  — ×2
Hard browser tasks (Browserbase hardest):   57% → 82%
Real cost of work: down ~25% on typical tasks,
  up to −45% on agentic ones — at the SAME list price
  (fewer tokens per task; cache reads cut by 75%)

What follows isn't my retelling — it's cases and techniques from the market, with authors and links. Check the primary sources yourself — I only collected and translated them into plain language.

00 Start

Market cases: who has already tested it — and what they measured

Three sources with numbers and names, not "experts say."

• Ramp (fintech) — the model ran ML experiments for 38 hours unattended, re-evaluating hypotheses as it went. VentureBeat's breakdown.
• Millennium (hedge fund) — Fable 5.1 traced a rare software crash the team had been investigating for years. Same source.
• CodeRabbit (Juan Pablo Flores and Gowtham Kishore Vijay) — ran 45 review tasks with 105 known issues: precision up 4.5 points, 186 fewer nitpick comments — the model learned restraint: it does what you asked and stops treating every blank space as a problem. And the surprise: the high reasoning setting was slower than low with no quality gain on their tasks. Full CodeRabbit report.

01 Marathon

Superpower 1: long tasks. Find your marathon

Which task has been sitting for months because "that's half a day of digging"?

Task: [TASK]
Output: [OUTPUT]

01 Marathon

Market technique #1: a goal instead of 22 steps

Ken Huang ("Agentic AI"): "your carefully engineered step-by-step prompts made output worse."

His live example: a 22-step migration prompt produced a worse result — including three wrong steps from the instruction itself. Replacing it with "goal + reason + constraints" gave a better result in fewer turns. The trick is the word "because": the reason measurably improves the micro-decisions the model makes on its own. Ken Huang's breakdown.

I'm working on this task: [TASK] — because I need: [WHY / OUTPUT].

I'm not spelling out the steps — plan them yourself. Take the task from start to finish, without stopping halfway and without asking permission for what I've already described.

Two requests along the way:
1. After each major stage — a 1–2 line update: what's done, what's next.
2. If data is missing — don't invent: ask once, with all questions collected together.

[attach materials or point to the folder/files]

The safety lines come from Anthropic's official Fable 5.1 prompting guide: the model sometimes works quietly (asking for updates fixes it) and may stop early to ask permission for work you already requested ("take it to the finish" fixes that).

02 The effort dial

Superpower 2: one dial instead of prompt gymnastics

Effort is how long the model "thinks." Five settings: low, medium, high, xhigh, max.

Anthropic's official position: effort is the primary control for quality, speed and cost on Fable 5.1. Picking the right level matters more than almost any prompt optimization.

Start: high (the default). Then test downward:

medium → roughly Fable 5's top quality, at lower cost.
         The workhorse for everyday tasks.
low    → often beats smaller models on cost per task
         while scoring higher. For the simple and the bulk.
xhigh/max → only for the hardest problems: the model may
         think for a long while before a long answer — normal.

Low-effort quirk: the model searches the web less often
on its own — ask for search explicitly when freshness matters.

In plain words: the "expensive top model" and the "cheap fast one" are now the same model with a dial. Don't switch models — turn the dial. The cheat sheet is assembled from Anthropic's guide and Ken Huang's observation that low-effort Fable often beats previous models at their max — the economics flipped.

03 Writing

Superpower 3 (with a catch): better writing — but denser

Fewer stock phrases and less jargon. But longer sentences, fewer paragraph breaks.

From Anthropic's official guide: Fable 5.1's writing is "a step up from earlier models, with fewer stock phrases." My experience agrees. The same doc describes two side effects — both cured by a line in the prompt (the fixes below adapt their wording):

Please remove all mannered prose: no decorative metaphors, no phrases that exist to display the writer. When a literal phrase is available, use it. Short sentences, frequent paragraph breaks.

Use lists and headers where the content is multifaceted — don't shy away from structure. (Unlike older models, this one under-uses bullets rather than over-using them — ask for structure explicitly.)

And audit your old prompts: if they still carry "no bold, no bullets" rules (we all wrote those against older models) — remove them. On Fable 5.1 they make answers poorer than they need to be.

03 Writing

Small things worth knowing

Three facts with no action required — they'll come up in conversations.

• Fewer false refusals. Anthropic reports ~60% fewer safeguard triggers on harmless requests (measured in Claude Code). If it still refuses something benign — rephrase positively: not "does this compile without errors?" but "are there any bugs in this program?"
• Invisible watermarks. Fable 5.1 text carries invisible marking (an EU AI Act requirement for models released after Aug 2, 2026); a detection API is in private preview for regulators and media. The practical takeaway is unchanged: the final pass is yours — your text should be yours in substance, not just in signature.
• Fable vs Mythos. Same underlying model: Mythos 5.1 is the relaxed-safeguards version for vetted organizations (cybersecurity, life sciences). You want Fable.

04 Finale

Should you switch? Test on your task, not on reviews

Where to next

• The switching rule: same task, same prompt, effort medium — compare against your current model on your own criteria. 15 minutes.
• Sources and authors: Anthropic's announcement · official Fable 5.1 prompting guide · Ken Huang — intent-based framing · CodeRabbit — measurements on 45 tasks · VentureBeat — pricing and the Ramp/Millennium cases
• Want prompting techniques with receipts? — the Prompt Engineering Field Guide