Why does Claude Fable 5.1 write differently from the model you were using last month?
Claude Fable 5.1, released by Anthropic on 1 September 2026, writes denser prose, uses fewer bullets and bold, quotes source documents without quotation marks, and answers from memory more often at low effort. None of these are bugs. They are documented behaviour changes, and each one has a short prompt fix published in Anthropic's official model guide.
Here is the honest limitation first. If you upgraded to Fable 5.1 and your summaries suddenly read like a wall of text, or your chat replies lost the structure you relied on, you did nothing wrong. The model simply has different defaults from Claude Fable 5 and Claude Opus 5.
Definition: Claude Fable 5.1 is Anthropic's current flagship model (API ID claude-fable-5-1). It has a 1M-token context window, 128k max output, adaptive thinking that is always on, and an effort setting (low, medium, high, xhigh, max) that controls how deeply it reasons before answering.
The good news is that every behaviour change on this list has a fix you can paste into your prompt or Project instructions in under a minute. Anthropic wrote them. I tested them on a marketing report, a meeting summary and a 40-page PDF. Below are the six that matter most if you are a non-developer using Claude in the app, in Projects, or through a no-code tool.
Source for everything in this article: Anthropic's What's new in Claude Fable 5.1 and Prompting Claude Fable 5.1 documentation, both current as of September 2026.
Fix 1: Which effort level should you use on Claude Fable 5.1?
Start at high, the default, then step down to medium for routine work. Anthropic states that Fable 5.1 at medium effort roughly matches Claude Fable 5 at lower cost, and that at low effort it is often competitive with Opus and Sonnet models on cost per task. Reserve xhigh and max for hard analysis where you have seen a real quality gain.
Effort is the single most important dial on this model. It replaces the old "extended thinking on or off" toggle: thinking is always on, and effort decides how much of it happens.
A practical rule that held up in my testing: drafting, rewriting, summarising and formatting work fine at medium. Anything involving a multi-step decision, a spreadsheet with formulas, or reading a dense chart deserves high. Only push to xhigh when you have compared the two outputs side by side and high actually fell short.
One important gotcha from the docs: at xhigh and max, the model may draft an entire long document inside its thinking and then write it out again as the reply. You wait longer and pay for the tokens twice. For long deliverables, stay at high unless you have measured a difference.
If you are on the API or a tool that exposes it, Fable 5.1 also lets you change effort mid-conversation without breaking the prompt cache (beta). Raise it for one hard step, lower it again for the next. Most practitioners will not touch this, but it explains why some tools now show an effort slider per message.
Fix 2: How do you stop Claude Fable 5.1 from writing dense, "mannered" prose?
Add one instruction to your message: "Please remove all mannered prose." Anthropic reports that Fable 5.1 sometimes writes longer sentences with fewer paragraph breaks than Fable 5, and that naming the anti-pattern (metaphor and flourish in place of direct statement) reliably shortens and clarifies the output.
This was the change I noticed first. Fable 5.1 is a better writer than earlier Claude models, with fewer stock phrases and less unexplained jargon, but it occasionally leans into style. A product description that should have been three plain sentences came back as one elegant paragraph nobody would skim.
Anthropic offers a long version of the fix that defines the problem, and a short version that usually works on its own. Here is the long one, copy-paste ready. Put it in the user message (preferred) or in your Project or system instructions:
Try this prompt:
Mannered prose substitutes metaphor and flourish for direct statement. Instead of "a parameter worth varying," the mannered writer produces "a dial worth turning." Instead of "this point still matters," they write "this point earns its keep." The phrases exist to display the writer, not to convey the idea, and readers can tell. That is why mannered prose irritates: it makes the reader work harder so the writer can perform. It is also imprecise. Metaphors drag in connotations the writer did not choose and cannot control. The fix is to say what you mean. When a literal phrase is available, use it.
In practice, on an internal newsletter draft, the long version produced visibly shorter sentences and more paragraph breaks, and the short version ("Please remove all mannered prose.") got most of the way there on its own. Start with the short one and escalate only if needed.
Fix 3: Why did Claude Fable 5.1 stop using bullets and bold in chat, and how do you get structure back?
Fable 5.1 uses headers, lists and bold less than earlier Claude models. If your saved instructions still contain anti-formatting rules written for older models ("never use bullet points"), they now over-suppress structure. Anthropic's fix is to delete those lines and replace them with a rule that states when formatting is appropriate.
This one catches almost everyone who built Projects in 2025. Back then, Claude over-formatted, so people added lines like "respond in prose only" or "no bold, no lists." Fable 5.1 already leans the other way, so those lines push it into paragraphs even when you asked for a comparison table.
Audit your Project instructions and custom instructions for anti-formatting language. Then replace it with Anthropic's recommended wording:
Try this prompt:
Use lists and bullet points when asked to, or when the content is multifaceted enough that they help with clarity. If the person explicitly requests minimal formatting, always format your responses without bullet points, headers, lists, or bold emphasis, as requested. In conversational, personal, or emotional exchanges, keep to plain prose.
The difference is that you are describing a decision rule rather than a blanket ban. The model then formats a five-option vendor comparison as a list and a one-line status update as a sentence, which is what you actually wanted all along.
Fix 4: How do you make Claude Fable 5.1 mark quotations when it summarises documents?
When summarising, Fable 5.1 is more likely than Fable 5 to reproduce passages from the source without marking them as quotations. Anthropic's fix is to add one complete worked example to your instructions: a sample request, a correct response, and a one-sentence rationale explaining why the response is correct.
For marketers and researchers this is the change with the highest risk. If you summarise a competitor's whitepaper or a news article and paste the result into your own deck, unmarked source sentences can slip through as if they were yours.
The fix works because Fable 5.1 follows examples closely. Rather than telling it "always mark quotes," you show it what a correct summary looks like: organised around agreement and difference, written in the model's own indirect speech, with at most one short marked phrase from the source.
Anthropic's example is built around two newspapers covering a bridge closure. Adapt the pattern to your own domain. The structure to keep is: the user request, the response, and a rationale line that states "each source is conveyed in one or two sentences of the assistant's own words; one short marked phrase; every other claim is reworded."
A lighter-weight version that worked in my testing for a single-document summary: "Summarise in your own words. If you keep any phrase of four or more consecutive words from the source, put it in quotation marks." It is less robust than the full example, but it is a one-liner you can add today.
Fix 5: How do you get Claude Fable 5.1 to finish the whole task instead of asking "Shall I continue?"
On long tasks, Fable 5.1 sometimes describes the next step ("Next, I'll…") or asks permission for work you already requested. Anthropic's fix is a short instruction telling the model you are not watching in real time, that reversible steps covered by the original request should proceed without asking, and that it should only stop for destructive actions or genuine scope changes.
This matters most when you hand Claude a multi-part job: "read these three PDFs, build a comparison table, then draft the recommendation email." Without the nudge, you may get the table and a polite question about whether to write the email.
Try this prompt (the opening sentence carries most of the effect):
You are operating autonomously. The user is not watching in real time and cannot answer questions mid-task, so asking "Want me to…?" or "Shall I…?" will block the work. For reversible actions that follow from the original request, proceed without asking. Stop only for destructive actions or genuine scope changes the user must decide. Offering follow-ups after the task is done is fine; asking permission before doing the work is not. Before ending your turn, check your last paragraph. If it is a plan, a question, or a promise about work you have not done, do that work now.
Two cautions from Anthropic. First, this block can make the model less likely to ask about ambiguous requests, so pair it with clear scope. Second, if you are working interactively and want to approve each step, do not use it. It is for delegated work, not pair work.
A companion instruction defines scope: "The user's request sets the scope, and the scope is the deliverable: don't quietly narrow, widen, or swap it. If one part is blocked, complete every other part and say exactly what you left out and why." Together they turn Fable 5.1 into a reliable finisher.
Fix 6: Why does Claude Fable 5.1 answer from memory at low effort, and how do you force a search?
At low effort, Fable 5.1 calls search and retrieval tools less often than Fable 5 and is more likely to answer from what it already knows. The simplest fix is to raise effort for that turn. The prompt fix is to tell the model that recognising a name is not the same as knowing its current state, and that such names must be searched as written.
This is the hallucination risk hiding inside a cost optimisation. Low effort is cheap and fast, which is exactly why you use it for quick lookups, and quick lookups are exactly where stale knowledge hurts. Ask about a tool's current pricing at low effort and you may get last quarter's numbers with total confidence.
Try this prompt:
When a query centres on a name you do not confidently recognise, or recognise from a fast-moving area like AI models and developer tools where the landscape shifts within months, the name itself is the thing to verify: search before answering, and include the name as the user wrote it in at least one query alongside any reformulations. This holds even when you have some background on it. Partial background is exactly what makes an out-of-date answer sound authoritative, so familiarity is not a reason to skip the search.
My rule of thumb: never use low effort for anything with a date, a price, a version number or a person's current job title. Medium plus this instruction is the floor for factual work. For a broader look at why staged rollouts mean your plan may not even have the newest model yet, see what a staged rollout is and why your paid AI plan got the new model late.
What should you actually change in your Claude setup this week?
Do three things: delete anti-formatting lines from your Project instructions and replace them with the decision rule in Fix 3; add "Please remove all mannered prose" to any writing template; and add the search-verification paragraph from Fix 6 to any Project that handles facts, prices or product names. Set effort to medium for drafting and high for analysis.
The bigger lesson is that model upgrades are not free. Every new version ships with different defaults, and prompts tuned for the last model quietly stop working. Anthropic now publishes a per-model prompting page for exactly this reason, and reading it on release day saves a week of confused iteration. If you want to build the fundamentals underneath these fixes, Anthropic's free interactive course is covered in our guide to Anthropic's official prompt tutorial.
Fable 5.1 is a better writer, a better researcher and a better long-task finisher than anything before it, once you tell it what you want. That is the whole game with AI at this level: it is not cold, it is just waiting for a clearer brief. We understand AI. We understand you better. With UD by your side, AI doesn't feel cold.
Reviewed by the UD AI team. Model behaviour and prompt wording verified against Anthropic's official Claude Fable 5.1 documentation on 9 September 2026. Effort names and defaults may change in future model versions.
Ready to turn these fixes into a workflow your whole team uses?
Knowing the six fixes is step one. Step two is building them into Projects, templates and AI staff roles that produce consistent output every day. We'll walk you through every step, from model and effort selection to prompt libraries and deployment, so the upgrade actually lands.
