Claude Fable 5.1 has five effort levels: low, medium, high, xhigh and max. They control how many tokens the model spends on thinking, tool calls and output. For most marketing writing, draft on low or medium and save high for the parts that need judgment. Keep xhigh and max for long research or strategy jobs where one bad assumption costs you more than the extra tokens.
Key Takeaways
- Fable 5.1 supports five effort settings. According to Anthropic's docs, the API default is
high, but claude.ai and Claude Cowork default to Medium (as of September 2026). - Effort affects every token: reasoning, tool calls and the final text. Lower effort gives you shorter, terser outputs with fewer tool calls.
- For marketing work, draft on low or medium, then use high on the headline, the argument and the offer.
- xhigh and max suit long research and positioning work. They're overkill for social captions or meta descriptions.
- At $10/$50 per million input/output tokens, most of the cost difference between effort levels comes from output and thinking tokens.
- Per-message effort (in beta) lets you lower effort partway through a conversation without losing your prompt cache.
What "effort" actually controls in Fable 5.1
In the API, effort is a single parameter, output_config.effort. Anthropic's effort documentation says it controls how many tokens Claude uses when it responds, and that it applies to all of them: visible text, tool calls and thinking. It isn't a creativity dial or a tone dial. It's closer to a budget for how hard the model works on the problem.
This matters for marketers because it changes more than "thinking time." On low, Fable writes tighter copy, makes fewer tool calls (fewer web searches, fewer file reads) and reaches a conclusion faster. On max, it'll go over the brief several times, test alternative angles and write longer answers unless you tell it not to.
Here's how Anthropic describes each level, as of September 2026:
| Level | Anthropic's description | What that means for a marketer |
|---|---|---|
low | Most efficient, significant token savings, some capability reduction | Captions, rewrites, variants, summaries |
medium | Balanced, moderate savings | Standard blog drafts, emails, briefs |
high | Default on most models; spends tokens as needed | Pillar posts, landing pages, argument-heavy content |
xhigh | Extended capability for long-horizon work | Multi-hour research agents, full content audits |
max | Absolute maximum, no token constraints | Positioning, pricing narratives, messy strategy calls |
Defaults differ by surface, so check where you're working
This is the part people get wrong most often. According to Anthropic's Fable 5.1 announcement, Fable 5.1 defaults to High effort in Claude Code and to Medium in Claude Cowork and Claude.ai. In the API, if you leave out the parameter, you get the model default, which the docs list as high for Fable 5.1.
So if your team compares "Claude in the app" with "Claude through our automation," it may be comparing Medium with High without realising it. Before anyone decides the API version "writes better," set effort explicitly on both sides.
A task-by-task effort map for marketing writing
I run content for edtech and startup clients, and this is the map I use now. Treat it as a starting point for your own tests. I haven't published benchmarks for it.
| Marketing task | Starting effort | Step up when |
|---|---|---|
| Social captions, hooks, 20 headline variants | low | Never; volume matters more than depth |
| Meta titles and descriptions at scale | low | The page is a money page |
| Rewriting a draft into brand voice | low or medium | The voice guide is long and subtle |
| Email newsletter from notes | medium | The email carries an offer or a price change |
| SEO / AEO blog post from a brief | medium | The topic is YMYL or technical |
| Landing page copy | high | Rarely go lower; conversion copy rewards judgment |
| Competitor teardown with web research | high | You need 30+ sources, so use xhigh |
| Quarterly content strategy or positioning doc | high → max | Stakes are high and inputs conflict |
The pattern: if volume is the job, go low. If judgment is the job, go high. Most blog content falls in between, which is why medium works well for it.
The "draft low, refine high" workflow
Nate Jones makes a useful argument in his piece on starting Fable 5.1 on low: with knowledge work, the hard part is deciding whether the work is right. A cheap first draft shows you the gaps quickly, and you then spend expensive tokens only where judgment is needed.
For marketing writing, I use this version of the workflow:
Step 1: Structure on low
Ask for the outline, H2s, the FAQ questions and a one-line claim for each section. It's quick and cheap, and it lets you see straight away whether the angle is wrong.
Step 2: Draft on medium
Once you've approved the outline, draft the full post on medium. For standard informational content, this is usually enough to get a publishable skeleton.
Step 3: Targeted high-effort passes
Don't regenerate the whole post on high. Paste in the three sections that matter, usually the intro and direct answer, the comparison or "how to choose" section, and the CTA, and ask Fable on high to tighten the argument, check the claims and find weak logic.
Step 4: Low for the last mile
Go back to low for alt text, meta description variants, social snippets and internal-link anchor suggestions.
This routine produces better work than running everything on high, because the high-effort passes are aimed at specific problems instead of spread across the whole post.
What effort costs, with real numbers
Fable 5.1 costs $10 per million input tokens and $50 per million output tokens, and cache reads cost $0.25 per million, according to Anthropic (as of September 2026). Output is five times the price of input, and higher effort mostly adds output and thinking tokens, so that's where the cost difference comes from.
A rough way to think about it, with illustrative numbers rather than measurements:
- A 2,000-word blog draft is about 2,700 output tokens of visible text.
- On low, total output (including any reasoning) may stay close to that.
- On high or max, reasoning and self-review can multiply output tokens several times over.
At $50 per million, even 20,000 output tokens costs $1. For a single post, effort barely affects cost. It starts to matter when you generate 500 programmatic pages or run agents all day. That's where low and medium pay for themselves.
Use caching and per-message effort together
Two features make effort settings much cheaper to use:
Prompt caching. Put your brand voice guide, ICP notes and style rules at the start of the prompt and cache them. Anthropic prices cache reads at $0.25 per million tokens, so a 20,000-token brand bible costs almost nothing to reuse on every request.
Per-message effort (beta). The effort docs describe a beta header, mid-conversation-output-config-2026-07-01, that lets you change effort partway through a conversation while keeping the cache. If you change the top-level effort value instead, you can lose the cache. With per-message effort you can do the strategy discussion on high and then drop to low for "now give me 15 LinkedIn hooks" in the same thread.
Where high effort earns its cost in marketing
In my own client work, going up to high or max has paid off in a few specific places:
- Positioning statements. When a client has three product lines and conflicting stakeholder opinions, higher effort gives you a better synthesis of the trade-offs. At lower effort you tend to get a bland average.
- Data-heavy posts. If the draft cites stats, higher effort does more checking and is more willing to say "I can't verify this." You still verify everything yourself.
- Answer engine optimisation (AEO) intros. The 2-3 sentence direct answer at the top of a post decides whether AI answer engines quote you. It's short, but it's worth the extra reasoning.
- Offer and pricing copy. Mistakes here cost real money, and more tokens are cheap by comparison.
When I scaled Masai School's LinkedIn from 50K to 160K followers, the volume work (hooks, carousels, repurposing) was never the bottleneck. The bottleneck was deciding which ideas to push.
Where high effort is wasted
- Short-form variants. Asking max for 30 Instagram hooks gets you slower, longer answers and no better hooks.
- Formatting jobs. Converting a transcript to a blog outline, turning tables into bullets, fixing heading hierarchy.
- Translation of approved copy. Unless you're adapting it for culture, medium is enough.
- Anything a human will rewrite anyway. If your editor rewrites 70% of the draft, you're paying for reasoning that gets deleted.
Prompting tips that interact with effort
Effort doesn't replace a good brief. A few habits that work better alongside it:
- Set length explicitly. Higher effort can produce longer answers. If you want 150 words, say "150 words max."
- Separate thinking from writing. On high, ask it to "decide the angle first, then write." You get a better angle and the copy stays clean.
- Give it a rubric. "Score this intro 1-5 on specificity, proof and clarity, then rewrite the weakest dimension." A rubric gives the extra reasoning something concrete to work on.
- Tell it what not to research. On xhigh, agents can spend a long time on tangents. Put limits on the sources and scope.
A simple team policy you can copy
Write the policy down:
- Default for all writing tasks: medium.
- Allowed to go low for anything under 300 words or any variant generation.
- High required for landing pages, pricing pages and anything with a stat in the headline.
- xhigh/max only with a named reason in the ticket (e.g. "quarterly positioning," "competitor audit").
- Log effort level in your content tracker for 30 days, then compare edit time per piece.
The last point is the one that counts. The right effort level for your team is the one that cuts human editing time, not the one that looks most impressive in a demo.
How Fable 5.1 effort compares with other models' settings
OpenAI's GPT-6 "Astra" has its own reasoning settings, and the names and defaults don't map one-to-one onto Anthropic's. Don't assume "high" means the same thing across vendors. If you're choosing between the two for a content team, test the same brief at comparable settings and measure edit time, not benchmark scores.
FAQ
What are the Claude Fable 5.1 effort levels?
There are five: low, medium, high, xhigh and max, set with output_config.effort in the API. Low is the most token-efficient and max has no token constraints. The API default for Fable 5.1 is high, according to Anthropic's docs as of September 2026.
What effort does Claude.ai use for Fable 5.1 by default?
According to Anthropic's launch post, Fable 5.1 defaults to Medium in Claude.ai and Claude Cowork and High in Claude Code. Keep this in mind when you compare app output with API output.
Is higher effort always better for blog posts?
No. For standard informational posts, medium usually gives a draft you can publish after editing. High helps with the intro, the argument and conversion sections, but running the whole post on max mostly makes it longer and slower.
Does effort change the writing style?
Not directly. It changes how much the model reasons and how verbose it tends to be, so it can indirectly affect length and depth. Tone and voice still come from your prompt and brand guide.
How much does max effort cost compared to low?
Pricing per token is the same ($10 input / $50 output per million as of September 2026). Max uses more tokens, mostly output and thinking tokens. For one article the difference is usually cents to a dollar. At scale it adds up.
Can I change effort in the middle of a conversation?
Yes, through a beta feature. Anthropic's docs describe per-message effort using the mid-conversation-output-config-2026-07-01 beta header, which keeps your prompt cache.
Which effort level is best for SEO meta descriptions?
Low. They're short and formula-driven, and you'll want many variants. Only step up for high-value money pages where the angle needs thought.
Should agencies set one effort level for all clients?
Set one default (medium works for most teams) and document when people may go up or down. Then check edit time per piece after a month and adjust.
Is Claude Mythos 5.1 better for marketing writing?
Mythos 5.1 isn't a realistic option for marketers. Anthropic restricts it to vetted organisations in its cyber and life-sciences verification programs. Fable 5.1 is the model to plan around.
Work With Me
I've spent 4+ years in organic growth, SEO and content, including growing Masai School's Instagram from 26K to 117K. The same test-and-measure approach applies to AI workflows like this one. If you want help setting up a content system where AI effort, human editing and distribution are balanced properly, have a look at my work and get in touch through the contact form at younusfardeen.in.