Claude Fable 5.2: What Anthropic Actually Says About Effort, Old Prompts, and Finishing the Job

On this page · 8 sections
- 01What Fable 5.2 is, and how it sits next to Opus 5.5
- 02Effort: start by knowing which default you are on
- 03It sometimes stops early. There is an official line for that
- 04Review old instruction files
- 05Three other official switches
- 06The 50% weekly cap, and what to measure first
- 07FAQ
- 08Want a second look at your coding-agent setup?
It is built for the kind of coding job that runs for hours, across a large repo, and needs to finish without you sitting on its shoulder.
A lot of new prompting advice started circulating the same week. Most of the useful parts were already in Anthropic's Fable 5.2 prompting guide, published with the model. This post walks through those official settings, in plain words, and flags the common misreadings.
We're AI Lab Australia, and we set up coding agents for teams. We have not run these settings on a client's repository. The figures below come from Anthropic's own documentation and its system card.
What Fable 5.2 is, and how it sits next to Opus 5.5
The model ID is claude-fable-5-2. On the public Claude API it costs US$8 per million input tokens and US$40 per million output tokens. Tokens are the chunks of text the model reads and writes. You pay for them, and a higher thinking setting usually means more of them. Fable 5.2 is not on Anthropic's Free plan.
Anthropic positions it for demanding, long-running work. On the product page it is blunt: for most everyday coding and knowledge work, start with Claude Opus 5.5.
That is worth taking seriously. A higher bill per token does not always mean a higher bill per job. On Cognition's FrontierCode 2.0 benchmark, which Anthropic reports in the Fable 5.2 system card, Fable at high effort scored 45.3% at an average of US$6.72 per task. Opus 5.5 at max scored 45.5% at US$12.82 per task. Similar score, about half the cost per task.
Two limits, both from the same card. These scores are vendor-reported, with no uncertainty intervals, so a gap of a few tenths of a point may not mean much. Fable's own top score on that test was higher still, at the xhigh setting, which we will come to.
If you already use Claude Code, Anthropic's coding app in the terminal, Fable 5.2 is a model you pick for the hard, long jobs. It is not the default model on any plan. Opus 5.5 is still the starting point for day-to-day work.
Effort: start by knowing which default you are on
Fable 5.2 has a setting called effort. It controls how much the model thinks before it answers, and how many tokens a turn uses. The levels are low, medium, high, xhigh and max.
The default depends on where you are sitting. On the public Claude API, effort starts at high. In Claude Code it starts at medium. If you open Claude Code, pick Fable 5.2, and leave the setting alone, you are on medium, not high. That single fact explains a lot of confused first impressions.
Anthropic's prompting guide says to start at high on the API, then move the setting up or down and test each change on your own tasks. It also says most of the extra skill already shows up at high. The step from medium to high raised scores on most of the coding and knowledge-work evaluations that Anthropic tracks. That step also uses more tokens and more time, so the same guide tells you to move back to medium wherever your own tests show that quality holds.
What each level is for, in Anthropic's words:
- low: quick back-and-forth, well-specified tasks, and high-volume work such as fast subagents.
- medium: interactive coding, chat, and everyday knowledge work.
- high: hard coding, research, agentic work (the model uses tools and works through a job in steps on its own), and long unattended runs. Turns take noticeably longer than at medium.
- xhigh: your hardest long-running work, where you have already measured a gain. On Anthropic's internal coding evaluations, it raised scores slightly and used many more tokens.
- max: for measuring the model's ceiling, or for a rare task where time and tokens don't matter. Turns are slow, and they use many more tokens than at xhigh.
A common shortcut is to park everything on high and treat medium as a waste. Anthropic still lists medium for everyday interactive work. If you are chatting through a change and reviewing each step, medium is the setting the guide describes for that job.
The other common shortcut is to treat max as off-limits. That is too sharp. On FrontierCode 2.0, Fable at max scored 45.7% at US$44.70 per task. That is barely above high (45.3% at US$6.72), and it sits below xhigh (49.1% at US$14.19). Max cost a great deal more than high, and scored lower than xhigh. On some other tests Anthropic reports, max can lead. So use max when you are measuring the ceiling, or when you have a rare task where the bill is beside the point. Don't make it the daily setting.
It sometimes stops early. There is an official line for that
On long, multi-step jobs, Fable 5.2 sometimes ends a turn before the work is done. The turn closes with a list of what remains, an offer to continue, or a stop at an ordinary obstacle such as a timeout. Anthropic says this happens more often at low and medium effort.
The official guide has a section called "Finish the whole task". It gives a block of text you can add to the system prompt. The gist is: keep going until the work is finished. Work through ordinary obstacles such as errors, timeouts, locked files and failing tools. Leave deliberate blockers alone, such as a file marked not to be touched. Treat that text as a starting point, and keep the distinction when you adapt it.
Anthropic replayed recorded sessions from an agentic coding product at medium effort. With a version of this text, about half as many sessions ended with doable work left. The text removed most endings that ask permission or offer to continue. Some endings that list what remains and then stop were still there. Output tokens rose by about 30 percent.
Because the model asks less often with this text, Anthropic says to keep your own confirmation step for risky or irreversible actions. That is the part we would keep even if you change the wording.
Review old instruction files
Fable 5.2 follows written rules closely. That is a virtue. It is less helpful when the rules were written for a different model.
Anthropic's guide has a section for people who carry over a system prompt, skills, agent files or a harness written for Claude Fable 5.1 or Claude Opus 5.5. The instruction is to review those files. In Anthropic's testing, some of those older lines raised cost, caused side effects, or no longer had an effect.
What to look for, from the official list:
- Standing instructions to delegate work to subagents. These can sit in system prompts, skills and agent files, and they can burn tokens.
- Lines that tell the model to think less, such as "answer directly without deliberating". Those often worked on Fable 5.1. In Anthropic's testing they did not stop Fable 5.2 from thinking at medium and above.
- Overly strict style rules. The model follows them closely, so a hard limit can make the prose choppy.
- Leftover cautions that have built up in memory files, notes and instruction files. The model might have written some of them itself.
A common misreading is to throw out every CLAUDE.md file and every skill. That is not what the guide says. It says review, trim the lines that no longer help, and keep a copy of the originals if you restart a system with its memory trimmed.
If you work in Claude Code, CLAUDE.md is the project instruction file the agent reads at the start of a session. Skills are extra instruction packs you can add for particular jobs. Both are worth a pass before you put a new model in front of them.
Three other official switches
Say who owns the code
Fable 5.2 responds closely to what the prompt says about who owns the code it is working on. In Anthropic's testing on small coding tasks, when the prompt said nothing about ownership, or presented the code as someone else's, the model tended to add to the existing structure, even when that repeated existing code. It also tended to report nearby problems instead of fixing them.
If you want it to improve the structure as it works, Anthropic suggests this sentence:
This codebase is yours: what it's for and how it's built are your call, and you'll keep building on it.
In the same testing, the model merged repeated code in about half of the tries with this sentence, and in under a tenth of the tries without it. The requested feature worked in nearly every try either way. The model also used about 1.5 times the output tokens with the sentence, and it made several changes nobody had asked for, many of them new tests and documentation. Sometimes it changed how an existing feature behaves. Anthropic's advice is to leave the sentence out where the user owns the code and wants it changed as little as possible.
A "don't think" line does not work at medium and above
At high effort and above, the model thinks before almost every reply, including a reply to a greeting. In Anthropic's testing, an instruction to answer directly without deliberating did not stop the model from thinking at medium and above. At low effort, the model can skip thinking on simple tasks.
If easy requests are taking too long or costing too much, lower the effort level for those turns. Do not add a "don't think" line and expect it to work at medium, high, xhigh or max. The cost follows the effort setting.
Tone and length come from the prompt
This one catches people out. Anthropic's guide is plain:
Effort barely changes how long the model's replies are. To change the length and tone of a reply, use your prompt.
If you want shorter answers, or a particular tone, put that in the prompt or in your instruction file. Turning effort up or down will not do that job.
One more official warning, kept short. Fable 5.2 can refuse a request that asks it to reproduce its internal reasoning in the response text. The refusal category is called reasoning_extraction. A related category, frontier_llm, covers requests that could help with the development of competing AI models. Anthropic's docs say not to ask the model to write its reasoning in the response. If you want to see how it reasoned, use the thinking display setting instead. A vague claim that any why-question will shut it down is too sharp. Asking why a function failed is ordinary debugging. Asking the model to dump its private chain of thought into the reply is what the classifier is built to catch.
The 50% weekly cap, and what to measure first
Fable 5.2 is available on all paid Claude plans. How it is billed depends on the plan, as set out in Anthropic's help article on Fable and your plan.
On Max and Ultra plans, and on premium seats on Team and seat-based Enterprise plans, you can use up to 50% of your weekly usage limit on Fable models at no extra cost. That is a share of the weekly limit you already have. It does not add 50% on top. Fable also uses the limit faster than other Claude models. When you hit the Fable share, you can keep going with usage credits, or switch to another model and stay inside the plan limit.
On Pro plans, standard Team seats and standard seats on seat-based Enterprise plans, Fable is not included in the plan's weekly usage limits. It runs on usage credits from the start, outside any limited-time launch offer.
Anthropic also says that, for a limited time after the Fable 5.2 launch, subscribers on eligible paid plans get extra Fable usage. We could not find a public end date or a published extra amount on the pages we used for this post, so we have not quoted one. Check the launch-offers note on the same help article if you are deciding whether to try it this week.
Before you change a setting, measure it on your own tasks. Anthropic's order is simple. Set effort to high, then run your tests at medium and xhigh as well. Read the instruction files you carry over. Run your longest agentic tasks and note early stops, extra approvals or repeated code. Then apply each official fix one at a time, and measure the effect and the token cost. A score on FrontierCode is interesting. It does not replace a test on your own repo, with your own files.
FAQ
01What is Claude Fable 5.2?
Anthropic's paid model for long, hard coding and agentic work, released on 9 November 2026. The model ID is claude-fable-5-2. List price on the public API is US$8 per million input tokens and US$40 per million output tokens. For everyday coding, Anthropic recommends starting with Opus 5.5.
02What effort level should I use for Fable 5.2?
On the API, start at high, which is the default. In Claude Code the default is medium, so raise it if you want the setting Anthropic uses as its API starting point. Then test medium, high and xhigh on your own tasks. Max is for measuring the ceiling, or for rare jobs where time and tokens are no object.
03Should I delete my CLAUDE.md file and skills for Fable 5.2?
No. Anthropic's guide says to review instructions written for earlier models and trim the lines that no longer help, such as standing subagent rules, "think less" lines, overly strict style rules and leftover cautions. Keep the files. Edit them.
04Does Fable 5.2 give me 50% extra weekly usage?
No. On Max, Ultra and premium Team and seat-based Enterprise seats, Fable can use up to 50% of the weekly limit you already have. It does not add an extra 50% on top. On Pro, standard Team and standard Enterprise seats, Fable runs on usage credits.
05Will a line that says "don't think" make Fable 5.2 cheaper?
Not at medium effort or above, in Anthropic's testing. At low effort the model can skip thinking on simple tasks. To spend less, lower the effort setting.
Want a second look at your coding-agent setup?
We help teams choose a model, set effort and instruction files, and keep a person in the loop for anything risky. If you already use Claude Code and you are trying Fable 5.2, we can go through your CLAUDE.md, skills and usage limits with you. Our Claude Code page explains the tool, and our AI models guide covers how the main model families compare.
If you want a straight answer on whether Fable 5.2 is worth the cap on your plan, email info@ailabaustralia.com with how you use Claude today. We'll tell you honestly, including if you should stay on Opus 5.5.

WRITTEN BY
Roman Silantev
CEO and Lead AI Engineer. Scopes and builds AI systems for Australian businesses.
← PREVIOUS6 n8n Google Workspace Workflows for Australian Small Businesses (Free JSON)
NEXT ARTICLE →AI Chatbot Cost in Australia: Per Month, Per Build, and What Actually Drives ItKeep reading in Tools & models
All articles →
Claude Code Skills: Five Public Packs, What They Actually Do, and What to Check First10 MIN READ
Ponytail for Claude Code: What 'Half the Tokens' Really Measures, and How to Install Agent Skills Safely10 MIN READ
Subframe Motion for Australian React Teams: 310 Free Animations Your Coding Agent Can Search9 MIN READ
