You switch your agent to Claude Fable 5.2, keep the same system prompt, and something feels off: easy questions take too long, long jobs stop halfway with an offer to continue, and the token bill grows. Anthropic has published a prompting guide for Fable 5.2 that explains most of these effects. The short version: several habits that helped earlier models now cost money or do nothing, and a few short sentences change the model’s behaviour more than you might expect.
Which effort level should you use?
Effort is the setting that decides how much the model thinks before it answers, how long a turn takes and how many tokens it uses. Fable 5.2 offers five levels: low, medium, high, xhigh and max. Anthropic recommends starting at high, which is the default on the Claude API, and then testing other levels against your own tasks.
The guide matches each level to a kind of work. Low suits quick back-and-forth, well-specified tasks and high-volume jobs such as fast subagents; in Anthropic’s testing, Fable 5.2 at low scored close to Fable 5.1 at higher effort. Medium fits interactive coding, chat and everyday knowledge work. High is for hard coding, research and long unattended runs. Xhigh is for your hardest long-running coding and agent work where you have measured a gain, and max is mainly for measuring the model’s ceiling or open-ended problems such as mathematics. Anthropic says most of the capability gain is already present at high, and above high the gains are uneven.
Why does the model stop before the job is done?
On long, multistep tasks, Fable 5.2 sometimes ends a turn early: it lists what remains, offers to continue, or stops at an ordinary obstacle such as a timeout. This happens more often at low and medium effort. The guide offers system-prompt text that tells the model to keep working through ordinary obstacles, such as errors or failing tools, but to respect deliberate blockers, such as a file marked not to be touched.
When Anthropic replayed recorded sessions from an agentic coding product at medium effort, a version of this text cut the number of sessions that ended with doable work left by about half. Output tokens rose by about 30 percent. Because the model asks for permission less often with this text, keep your own confirmation step for risky or irreversible actions.
What changes when you say the code belongs to the model?
Fable 5.2 reacts strongly to what the prompt says about who owns the code. When the prompt says nothing, or presents the code as someone else’s, the model tends to add to the existing structure, even if that repeats code, and to report nearby problems instead of fixing them. With one sentence saying the code base is the model’s own, Anthropic saw it merge repeated code in about half of the tries on small coding tasks, compared with under a tenth without it.
That sentence has a price. Output tokens rose to about 1.5 times, and the model made changes nobody asked for, often new tests and documentation. Its closing summary also left out small changes, so the guide advises reviewing the diff rather than only the summary.
Which old instructions should you remove?
Fable 5.2 follows instructions very closely, so rules written for Fable 5.1 or Opus 5.5 can now raise cost or have no effect. The clearest case is a line like “answer directly without deliberating”. At medium and above, such a line did not stop Fable 5.2 from thinking in Anthropic’s tests; the guide says to remove it and lower the effort level instead. Effort also barely changes how long replies are, so ask for the length you want in the prompt itself.
Other candidates for review are standing instructions to delegate to subagents, token budgets that a harness writes into the prompt, strict style rules that make prose choppy, and cautions that have piled up in memory files and notes. In one simulated test, a single old note saying “never retry” was enough to make the model retry a busy tool in only about four out of ten tries, down from nearly all of them.
How do you keep long agent runs visible?
During long runs of tool calls, Fable 5.2 can write very little text for the user. The updates it writes between tool calls arrive as progress-update thinking blocks, and with the default display setting their text is empty, so a client that only shows text can look silent. The guide’s first step is to request these updates with the “updates” display option. A short reminder that the harness appends after several silent steps roughly halved the longest silent stretch in Anthropic’s tests, with no noticeable rise in tokens or time.
On subagents, the guide is just as practical: check which lines in your prompts invite delegation before you change anything else, and check which model your subagents run on, because they can make up a large part of a session’s cost.
What happens if you ask for the model’s internal reasoning?
Fable 5.2 runs safety classifiers that can decline a request. One refusal category, reasoning_extraction, covers requests that ask the model to reproduce its internal reasoning in the answer. Asking which data an answer is based on is a normal question; asking for the hidden thought process is not. If you need to see how the model reasoned, the guide points to summarized thinking blocks instead.
Our assessment
The figures above are Anthropic’s own test results, often from small or simulated tasks, so treat them as a direction rather than a promise for your workload. Still, the pattern is useful: most of the advice is about removing text, not adding it. If you move a working agent to Fable 5.2, the cheapest first step is to set effort explicitly, delete the “think less” and “always delegate” lines, and only then add the finishing or ownership sentences, one at a time, while you watch the token count. Start with the job you need done; the effort level and the prompt should follow from it.
