Fable 5 is Back

Original Article ↗ Hacker News Discussion ↗

Anthropic’s high-end Claude Fable 5 AI model has been reinstated with a temporary promotion for subscribers, but with tight usage caps, aggressive safety guardrails, and plans to move it to pay-per-use pricing after July 7. Many users praise its planning and long-horizon coding abilities yet complain that safeguards frequently downgrade requests to older models, making it unreliable for security, ML, and even routine development work. The changes are fueling concern over opaque quotas, rising costs, and trust in US-based AI providers, prompting some to explore cheaper Chinese or open-weight alternatives and to question Anthropic’s strategy and communications.

Model capabilities and use cases

  • Many find Fable 5 noticeably better than Opus 4.8 for planning, long-horizon work, complex refactors, and orchestrating sub-agents; some say it doubles their development speed or handles tasks they’d pay a human five figures for.
  • Others report only modest gains over Opus or even GPT‑5.5 for everyday coding, math, or prototyping, saying Sonnet‑class or cheaper models are “good enough” for most work.
  • Fable/Claude Design gets specific praise for UI/UX design, redesigning screens, and converting plain HTML into modern frameworks (React/Nuxt/Angular), with some small SaaS owners using it heavily for product evolution.

Pricing, subscriptions, and token economics

  • Temporary promo: until July 7, users can spend up to 50% of their weekly subscription limits on Fable 5; after that it’s usage‑credits/API priced only.
  • Many see this as a “first hit is free” upsell and worry subscriptions will stagnate on older models (e.g., Opus 4.8) while new SOTA is pay‑per‑token only.
  • Users report burning through Fable limits or 5‑hour sessions in minutes due to aggressive parallel workflows, making it feel unusable for continuous work.
  • There’s frustration that subscription “usage” bars hide actual token numbers, change over time, and are hard to audit; some track costs via third‑party tools and conclude Pro/Max are heavily subsidized loss leaders.
  • Several say if Fable is only API‑priced at high $/Mtok, they’ll shift most workload to cheaper providers or open‑source, keeping Anthropic only as an occasional “planner.”

Safety guardrails, downgrades, and usability

  • A central complaint: Fable’s safety classifier frequently flags benign work (coding, cybersecurity, DSP, Kubernetes, even vitamin stacks or book drafts), then silently or explicitly downgrades to Opus 4.8.
  • Users see background tasks spawned by Fable that then get blocked by the safeguards Fable itself triggered, killing long‑running workflows.
  • Guardrails are perceived as much stricter after the U.S. export‑control episode; some note it now blocks even basic security review or login automation that previously worked.
  • People resent paying for tokens consumed by failed/flagged calls and by broad, sometimes regex‑like classifiers, and say this undermines Fable’s advertised “tenacity” and agentic autonomy.

Trust, governance, and geopolitics

  • Several argue the export‑control saga and Anthropic’s messaging (“too powerful,” “frontier risk”) damaged trust: companies fear building on a model that regulators can pull overnight.
  • Others defend cautious rollout, citing documented Mythos/Fable cybersecurity findings (e.g., large volumes of Firefox bugs) as evidence risks are real.
  • There is ongoing anger over earlier issues: secret downgrades on ML topics, hidden fingerprinting, mandatory data retention for Fable, and proposals seen as hostile to open‑weight models.
  • Opinions on Anthropic’s ethics diverge: some call it among the most trustworthy labs; others say it’s adversarial and paternalistic, no better than competitors.

Competition and alternatives

  • Many users actively compare Fable/Opus with GPT‑5.5/5.6, GLM‑5.2, DeepSeek, Grok, Composer, and local/open‑weight models.
  • GLM‑5.2 is frequently cited as “good enough” or Sonnet‑level for a fraction of the price, especially via cheap harnesses like OpenCode, though some say it struggles on very complex, non‑web tasks.
  • A recurring theme: if U.S. models become too expensive or locked‑down, users will route more work to Chinese/open models and keep U.S. frontier models only for planning or niche tasks.

Overall sentiment and outlook

  • Enthusiasm: Fable 5 is seen by many as a genuine capability jump for planning and large refactors; some small SaaS owners say it’s transformative despite cost.
  • Skepticism/resentment: equally strong pushback on pricing, opaque quotas, short promo window, guardrail overreach, and confusing comms (“banned for being too good,” then heavily nerfed).
  • Several expect market pressure and competition to eventually force Anthropic to include stronger models in subscriptions again or risk losing users to OpenAI and open‑weight ecosystems.