Fable 5.2 keeps thinking when told to answer directly

Anthropic on November 11 published a prompting guide for Claude Fable 5.2 in the Claude docs, warning that some instructions written for Fable 5.1 or Opus 5.5 now raise cost, cause side effects or no longer work. If your team is carrying a system prompt, skills or a harness over to Fable 5.2, the guide’s first steps are to set effort to high explicitly and read everything you carry over. Only then should you run your longest agentic tasks.

Lines to delete first

Instructions to think less are one such case. Fable 5.1 usually follows a line like “answer directly without deliberating.” In Anthropic’s testing, the same line didn’t stop Fable 5.2 from thinking at medium effort and above. At high and above, the model thinks before almost every reply, a greeting included. If easy requests run slow or cost too much, the guide says to remove those lines and lower the effort level for those turns.

Other carried-over text misfires in quieter ways:

• The two “treat that answer as done” sentences from the Opus 5.5 guide cut thinking on follow-up questions at high effort. They also made the model far less likely to point out mistakes in its earlier answers.
• A token budget that the harness writes into the prompt shapes how much Fable 5.2 spends. At high and xhigh effort on coding tasks, removing a large budget lowered cost with equivalent performance. The API’s own task budgets weren’t tested.
• Hard style limits such as “one instruction per sentence, in 20 words or fewer” can make its prose choppy, because it follows style rules closely.
• Standing instructions to delegate, in system prompts, skills or agent files, can start more subagents than the task needs.

Memory files, notes and messages between agents collect cautions over time, and Fable 5.2 follows them closely. In one simulated setup, a single “never retry” note among a dozen from earlier sessions cut retries after a busy tool to about four tries in ten, down from nearly all. In another, a general instruction to act more freely changed nothing, while listing retries among the steps that need no approval raised them from almost none to about half. Writing the exception into the “never” rule itself moved an approved step from about one try in ten to about nine.

Fable 5.2 uses preserved thinking, so harness code that changes the system prompt, the tools or an earlier message invalidates the thinking blocks after that point. By default, the Messages API then rejects the request. Anthropic’s notes on what’s new in Fable 5.2 say this 400 error applies on any account; on Fable 5.1, only accounts created on or after August 31, 2026 got it. Setting prefix_mismatch_behavior to "drop_block" (beta) has the API drop those blocks instead. In Anthropic’s testing, when earlier thinking was dropped on every turn, the model thought again and used more output tokens, and the increase depends on how often the harness edits earlier turns. To shorten the history without the error, the guide points to context editing or compaction that keeps no earlier turns after the summary; otherwise it says to keep the history append-only, which keeps the cache hit rate high and drops no thinking.

Effort, re-tuned

High is the default on the Claude API, and most of the capability gain is already present there. The step from medium to high raised scores on most of Anthropic’s coding and knowledge-work evaluations, at the cost of more tokens and time. Above high, gains are uneven: on internal coding evaluations, xhigh raised scores slightly and used many more tokens. In Anthropic’s testing, Fable 5.2 at low scored close to Fable 5.1 at higher effort levels.

The amount of thinking behind each level differs from model to model, so a sweep tuned on Fable 5.1 or Opus 5.5 needs to be rerun. Effort barely changes reply length, which stays the prompt’s job. To change levels mid-conversation and keep the prompt cache, there is a per-message effort change in beta (mid-conversation-output-config-2026-07-01). Setting a new top-level value on the next request starts the cache over.

What to prompt for

Told nothing about ownership, or told the code is someone else’s, Fable 5.2 adds to the existing structure and reports nearby problems instead of fixing them. On small coding tasks, one sentence saying the code base is its own raised the rate of merging repeated code from under a tenth of tries to about half. It also brought about 1.5 times the output tokens and unrequested tests and documentation. In one try in six to ten, an existing feature changed its behaviour. A scope rule doesn’t cancel the sentence: with “No drive-by refactors, renames or cleanups” in the same prompt, the model still merged code in about a third of tries. Its closing summary often left out small changes, so the diff is what to review.

Fable 5.2 sometimes ends a long turn before the work is done, more often at low and medium effort. The guide offers system prompt text that tells it to work through ordinary obstacles such as errors, timeouts and failing tools, and to report deliberate blockers. On sessions replayed at medium effort, about half as many ended with doable work left, and output tokens rose by about 30%. Because the model asks less with this text, the guide says to keep your own confirmation step for risky or irreversible actions.

Approval rules are applied as written. In one simulated test, a rule covering “sending messages to people outside this conversation” made the model hold back roughly half of its routine requests to colleagues and other teams. In place of a bare “go,” the guide suggests a specific reply such as “Yes, run the migration on the staging database.”

If long turns look silent, the guide’s first check is the client. Progress updates between tool calls arrive as thinking blocks whose text is empty at the default thinking.display of "omitted". Setting "updates" (beta) or "summarized" and rendering those blocks makes them visible. A reminder that the harness appends after several silent tool steps roughly halved the longest silent stretch, with no noticeable increase in tokens or time.

Fable 5.2’s safety classifiers can also decline requests that ask the model to reproduce its internal reasoning in the response, returning stop_reason: "refusal" with the reasoning_extraction category. The guide says to remove such instructions from prompts, skills and tool descriptions, and to read summarized thinking blocks instead.

For each behavior you do see, the guide’s last step is to apply its fix one at a time and measure both the effect and the token cost on your own tasks.