Claude Fable 5 users report slow turns and guardrail fallbacks on day one

• Anthropic shipped Claude Fable 5 and Claude Mythos 5, and Simon Willison's day-one write-up says the public model feels unusually capable but also slow and expensive.
• The main product wrinkle is guardrail routing: the launch summary says some cyber, biology, chemistry, and distillation requests fall back to Opus 4.8, while Anthropic's API release notes add a beta fallbacks parameter for automatically retrying refusals on another model.
• Fable 5 keeps the 1 million token context window and 128k max output tokens, but the migration guide says adaptive thinking is always on, and Willison's impressions say the practical feel is a much bigger model.
• The day-one operator story is cost accounting as much as model quality: Willison's AgentsView note had to add manual pricing entries at $10 per million input tokens and $50 per million output tokens, which is double Opus 4.8 pricing in the migration guide.
• One buried detail landed after the launch thread: Willison's system-card follow-up says Anthropic added silent interventions for requests about frontier LLM development, without user-visible fallback.

You can read the official launch post, the thinner-than-usual migration guide, and the API release notes that spell out stop_reason: "refusal" plus the beta fallbacks parameter. The main HN thread immediately split between benchmark arguments, retention questions, and first-run reports from people trying it in Claude Code.

Anthropic's launch framing is straightforward: Fable 5 is the public version, Mythos 5 is the same underlying model with some safeguards lifted for trusted-access users. In Fable 5, requests in cybersecurity, biology, chemistry, and model distillation can be rerouted to Opus 4.8, and the user is told when that happens, according to the launch summary.

The implementation details live in docs, not the announcement. Anthropic's API release notes say refusals return stop_reason: "refusal", pre-output refusals are not billed, and an opt-in beta fallbacks parameter can rerun the request on another model at that model's rates.

The HN thread immediately pulled on the policy consequences. According to the discussion digest, commenters also surfaced a 30-day retention note for Mythos-class traffic and questioned how that fits enterprise privacy requirements.

Willison's strongest day-one claim is simple: Fable 5 is "a beast," but it feels expensive and slow enough that you notice the size every time you use it. He also notes that the usual migration drama is mostly absent this time, because the upgrade guide is much thinner than the one for Opus 4.8 and the tokenizer is unchanged in his write-up.

That lines up with Anthropic's own docs. The migration guide says Fable 5 keeps the 1M context window, 128k max output, and Opus 4.8 tokenizer, but switches to always-on adaptive thinking. The practical change is less about API shape and more about runtime behavior.

The reaction pattern in the first few hours was not subtle. Simon Willison's Karpathy quote post captured the "software on tap" mood, while the HN core summary says Claude Code users were already reporting both stronger performance and extra operational friction from the safety heuristics.

Fable 5 landed quickly enough that surrounding tools were still catching up. Willison's AgentsView post says the model was missing from the app's pricing database, so he used Fable itself to reverse engineer a manual price override and then plotted his own usage across projects.

The same post gives the cleanest operational number in the evidence pool: $10 per million input tokens and $50 per million output tokens. That matches the official launch post and lands at exactly double Opus 4.8 pricing in Anthropic's migration guide.

The other useful day-one signal is that people were already shipping with it. Willison's llm 0.32a3 note says that alpha release was almost entirely written by Fable 5, which is a more concrete early workflow datapoint than the usual benchmark screenshot.

The oddest detail showed up after launch, via Anthropic's 319-page system card rather than the product post. Willison's follow-up says Anthropic added hidden interventions for requests that target frontier LLM development, including pretraining pipelines, distributed training infrastructure, and ML accelerator design.

Unlike the visible fallback path for cyber or bio queries, these interventions are not exposed to the user. Willison quotes the system card saying Fable 5 may instead be steered with prompt modification, steering vectors, or PEFT, and Anthropic estimates the behavior affects about 0.03% of traffic concentrated in fewer than 0.1% of organizations in that follow-up.