Claude Fable 5.1: Seven Prompting Changes That

Anthropic's Claude Fable 5.1 finishes more of the work you give it, and it does that work for less money than its predecessor. Those two sentences sound like an unambiguous upgrade. They are, mostly. But a new video from the AI Labs channel argues that the model behaves differently enough from Fable 5 that prompting it the old way can leave you with worse results and a bigger bill.

The video, "7 Rules To Use Claude Fable 5.1 Better Than 90% Of The People," runs about twelve minutes and is built on what the team says is real client work rather than benchmarks. I went through it, plus the surrounding evidence, and the picture that emerges is more interesting than the tips themselves: Fable 5.1 is a model optimized for endurance, and endurance changes everything about how you should use it.

The Claim Versus the Tests

Anthropic marketed Fable 5.1 as better at finishing tasks. The AI Labs team pushed back on the framing in their video: "Fable 5.1 isn't doing better work. It's getting through more work for less money."

Their evidence: they gave Fable 5 and Fable 5.1 the same tasks on their own client projects, ran each task multiple times, and compared. Fable 5 did better work on most of the tasks. Fable 5.1 only won on the hardest one, where it finished around 40% faster and cost well under half as much. On the tasks where its work wasn't better, it still finished every attempt, while Fable 5 stopped before the end on some runs.

That pattern matches the economics. Models don't carry a conversation from one message to the next; the whole context gets resent with every prompt. Claude mitigates this with cached reads, saved copies of already-processed conversation reused at a discount, and Anthropic cut that cached-read price by 75% for Fable 5.1. The discount compounds on long tasks, where the same growing conversation is resent again and again. By Claude Code's own usage estimate, AI Labs found Fable 5.1 around 47% cheaper per session.

The catch sits on the output side. Artificial Analysis, the independent model-testing outfit, found Fable 5.1 generates roughly 1.7 times more output than Fable 5, which pushed their per-task cost up about 20% despite the cheaper reads. Output pricing didn't drop. A model that writes more costs more, even when reading got cheaper.

The Guardrail Problem Nobody Asked About

The strangest finding in the video is about what happens when the model hits a safety guardrail. Anthropic says guardrails interrupt Claude Code sessions around 60% less often with Fable 5.1. But when a guardrail does trigger, Claude Code can now switch to a weaker model without telling you and keep going. That weaker model then handles the rest of the session, even for requests unrelated to whatever triggered the switch.

The failure case AI Labs describes is almost comedic: a developer building a game had a joke file containing the word "biological." That single word was enough to trigger a switch. The project had nothing to do with biology. Instead of a refusal, which at least announces itself, the developer got quieter, weaker work and assumed Fable 5.1 produced it.

If your project legitimately touches a guarded subject, security research, biology, anything adjacent, the video's position is blunt: no setting and no phrasing avoids it. That's a design choice worth scrutinizing. A visible refusal is annoying but honest; a silent downgrade is a correctness problem disguised as a usability improvement. Anthropic presumably sees fewer interruptions as the win condition. For anyone doing serious work, the win condition is knowing what model answered your prompt.

Hallucinations Went Up Too

Artificial Analysis scores Fable 5.1 higher than Fable 5 on accuracy, but the hallucination rate rose. When Fable 5.1 didn't know an answer, it invented one 72.6% of the time, against 63.6% for Fable 5. The video's take is the right one: a model that says "I don't know" is more useful than one that confabulates with confidence, because you can verify an admitted gap. An invented answer presented as fact goes straight into your work and only surfaces later, usually at the worst possible moment.

The Seven Tips, Compressed

The practical advice follows from the behavior. The highlights:

Turn effort down. Code Rabbit, which builds automated code review, tested Fable 5.1 on real review work and found low effort caught 61% of problems versus 57.1% on high, finishing about three minutes faster. The AI Labs team saw the same on their own project: high effort made the model loop on one part of a feature, while low effort shipped the whole thing better. Their rule: start low, raise it only when the result demands it.
Delete your old instructions. If you spent months patching Fable 5's quirks in your CLAUDE.md, those patches now constrain a model that doesn't have the quirks. Test removals one at a time; the video also points to a Claude Code prompt-audit command that flags stale instructions for review.
Tell it nobody's watching. Fable 5.1 pauses for permission on steps already covered by your request, which means a long unattended job stalls the first time it gets cautious. Anthropic's own prompting guide recommends stating that no one is monitoring and that it should proceed with anything reversible already covered by the brief.
Say what to leave alone. Ask for one feature and the model may fix neighboring problems or edit your test files uninvited. Scope the task, and instruct it to report unrelated issues rather than fix them.
Demand edits, not rewrites. Change one function and it'll rewrite the entire file, burning output tokens on identical lines. Given the 1.7x output inflation, this habit compounds fast.
Kill the mannered prose. Anthropic's term for the model's fondness for phrases like "a dial worth turning" instead of "worth changing." Fable 5.1 uses less of it than Opus 5 or Fable 5, but its sentences run longer with fewer breaks. The four-word fix, per the video: instruct it to remove all mannered prose. (As someone whose entire job is prose, I feel seen and slightly threatened. lmao.)
Give it the big job. Don't slice a feature into small prompts. Fable 5.1's edge is completing long work without stopping, so the whole-feature prompt is where it beats its predecessor. We've covered adjacent territory before in Claude Fable 5 Prompting Habits That Actually Matter, and the throughline holds: agent models reward clear, complete briefs over hand-holding.

What the Video Doesn't Settle

Two caveats keep me from treating all this as settled. First, the headline comparisons come from one team's client work, with project details withheld, so the sample is unauditable. Second, and more fundamentally, the whole playbook assumes you're on Anthropic's rails. BleepingComputer reported that Anthropic kept Fable 5 free for paid users until July 19 as it bought time during rollout, a sign the company itself treats model transitions as disruptive enough to buffer. Ecosystem alternatives keep multiplying too; Playcode's roundup of Fable 5 alongside Sol, Grok 4.5, and GLM 5.2 is one snapshot of how fast the field moves, and community repos like the Claude Fable 5 Having Fun project show users stress-testing these models well outside office benchmarks.

The deeper tension is the silent fallback. Every other tip assumes you know which model is working for you and can tune your prompts accordingly. A system that swaps in a weaker model mid-session makes that assumption false, and no amount of prompt hygiene fixes a transparency problem.

Still, the practical core of the video survives scrutiny: Fable 5.1 is a marathoner, price-tuned for long sessions, chattier than its predecessor, and more willing to fill silence with invention. Treat it like a sprinter and you'll pay for the mismatch. Treat it like a marathoner with a tendency to embellish, and you'll get your money's worth.

Yuki Okonkwo covers AI and machine learning for Buzzrag.