Fable 5 is Anthropic's strongest public model, and it quietly hands your request to Opus 4.8 when a safety classifier fires. Here is what triggers the fallback, what the community found in week one, and the prompt changes that keep you on the model you are paying for.

Anthropic shipped Claude Fable 5 on June 9, and within 48 hours the launch thread on Hacker News passed 2,500 points. The reaction splits cleanly down the middle. Half the discussion is people watching it do things no public model has done. The other half is people asking why it refused to edit a resume.

Both halves are real. Fable 5 is the strongest model you can rent right now, and it ships with safety classifiers that sometimes hand your request to a different model entirely. If your prompts were written for Opus 4.8 or anything older, some of them now actively work against you. Here is what the model does, what trips the classifiers, and how to write prompts that stay on Fable.

What Fable 5 actually is

Fable 5 is Anthropic's Mythos-class model with safeguards bolted on. The same underlying model ships as Claude Mythos 5 to a small set of vetted partners with the safeguards lifted. Everyone else gets Fable: same weights, plus classifiers watching the conversation.

The capability jump is the largest Anthropic has shipped. It posted the top score on Cognition's frontier coding eval and was the first model to break 90% on a long-running analytics benchmark, a 10-point jump over Opus 4.8. Stripe, an early tester, reported it migrated a 50-million-line Ruby codebase in a single day, work they had scoped at two months. The vision upgrade is large enough that it plays Pokémon FireRed from raw screenshots, where earlier models needed helper tools to parse the screen.

The headline shift for everyday use is duration. Fable 5 is built for tasks that run hours or days inside an agent harness. Single requests at high effort can run many minutes while it gathers context and verifies its own work. It costs $10 per million input tokens and $50 per million output, double Opus 4.8, and API traffic now carries a 30-day data retention requirement.

Why it refuses, and where your request actually goes

Fable 5 runs safety classifiers targeting offensive cybersecurity, biology and chemistry, and attempts to extract its internal reasoning. When one fires, your request does not error out. It gets answered by Claude Opus 4.8 instead, and you get a notice that the handoff happened. On the API the response includes a reason; in agent harnesses you can wire an automatic fallback or handle the refusal stop reason yourself.

Anthropic says the classifiers fire in under 5% of sessions. Week one suggested the false positives cluster painfully. The Register documented accounts where the input classifier fired on the literal word "hello," a user who could not get help editing an "Application Security Architect" resume because the title pattern-matched to security tooling, and an immunologist whose work kept tripping the biology classifier on the word "cancer." Anthropic acknowledged the tuning was too stringent, called it the wrong tradeoff, and committed to surfacing refusal reasons and loosening the biology triggers for legitimate research.

The practical consequence: when Fable 5 gives you a noticeably weaker answer, check whether you got Fable at all. The Opus fallback is a good model, and that is exactly why a silent handoff is easy to miss. If your output quality dropped mid-session, look for the fallback notice before you blame your prompt.

The community verdict so far

Reddit reads cautiously impressed and fixated on the boundary. The most upvoted complaint frames the launch as a preview of AI inequality, since the unrestricted Mythos 5 goes only to Project Glasswing partners and select biology researchers, while the public gets the filtered version. Security researchers are the loudest critics for a concrete reason: the cyber classifier blocks defensive work alongside offensive work, and Anthropic's own docs admit benign security tasks can trigger it.

Builders on X lean enthusiastic, and the pattern in their reports matches Anthropic's guidance: the people getting the most out of Fable 5 are handing it their hardest unsolved problems. Testing it on the same small tasks you gave Opus undersells it, because the gains concentrate in long, ambiguous, multi-step work.

The prompt fixes

Fable 5 changes what good prompting looks like. These are the adjustments that matter, drawn from Anthropic's own migration guide and from what early users hit in week one.

Stop asking it to show its reasoning. This is the one that catches working prompts. Any instruction telling the model to echo, transcribe, or explain its internal reasoning in the response can trigger the reasoning-extraction classifier and dump you to Opus. "Think step by step and show your work" was harmless boilerplate for years. On Fable 5 it is a fallback risk. If you need visibility, read the summarized thinking blocks the API provides instead of asking for reasoning in the response text.

Old: Think through this step by step, explain your reasoning, then answer.

New: Work through this carefully, then give the answer with the key evidence for it.

Delete the scaffolding you wrote for weaker models. Prompts that enumerate micro-steps, dictate the order of operations, and pre-chew the reasoning were compensation for models that needed it. On Fable 5 they degrade output, because the model follows your weaker plan instead of building a better one. State the goal and the success criteria, and let it plan.

Migrate this service to the new auth library. Done means: all tests pass, no remaining imports of the old library, and the README install section reflects the change.

Add context for anything near a classifier. Borderline topics survive when the request carries its own legitimacy. The resume case fails as "Application Security Architect resume" and works when the framing makes the intent unambiguous.

I am a job seeker updating my resume. Rewrite my work history bullets to be more concrete. The role title is Application Security Architect.

Rein in the elaboration. At high effort Fable 5 surveys options it will not pursue and explains root causes at length. One short instruction fixes it without listing every verbose habit by name, because instruction-following is strong enough now that brief steering works.

Lead with the outcome. Drop details that do not change what I would do next.

Pick an effort level on purpose. Effort is the main lever for cost and latency, and low effort on Fable 5 often beats max effort on prior models. Default to high, drop to medium or low for routine work, and reserve xhigh for the tasks where capability is the whole point. If a simple task is taking minutes, you are paying deliberation tax on work that did not need it.

Tell it why, along with what. Fable 5 measurably improves when it knows the intent behind a request, because it connects the task to relevant context instead of guessing.

I am preparing a pricing page for a product that has not launched. The copy needs to make the mechanism clear without claiming results we do not have. With that in mind: review this draft.

Check your prompts before you migrate

The failure mode with Fable 5 is quiet. A prompt that worked on Opus either silently underperforms because its old scaffolding constrains the model, or silently triggers a fallback and gets answered by Opus anyway. Either way you pay Fable prices for sub-Fable results.

That audit is what FixMyPrompt does. Paste a prompt you plan to run on Fable 5 and the QA flags the patterns this model punishes: show-your-reasoning instructions that risk the extraction classifier, step-by-step scaffolding written for weaker models, missing intent that leaves borderline requests unframed, and absent success criteria on long-running tasks. The rewrite restructures around goal and criteria, which is the shape Fable 5 rewards. Three free checks a day, no signup.

The short version

Fable 5 is Anthropic's Mythos-class model with safety classifiers in front, and when one fires your request gets answered by Opus 4.8 with a notice. The classifiers target offensive security, biology, and reasoning extraction, and week one showed false positives on harmless requests. The biggest prompt changes: never ask it to show its reasoning in the response, strip out micro-step scaffolding written for older models, frame borderline topics with explicit intent, and steer length with one short instruction. Hand it harder problems than you gave Opus. That is where the price difference earns itself back.

Claude Fable 5: What It Can Do, Why It Refuses, and the Prompt Fixes That Stop the Opus Fallback

What Fable 5 actually is

Why it refuses, and where your request actually goes

The community verdict so far

The prompt fixes

Check your prompts before you migrate

The short version

Sources

Read next

Run a free QA on your own prompt