Anthropic's Claude Fable 5 is the best publicly available model, but comes with strict guardrails and higher cost

Anthropic's Claude Fable 5 represents a genuine step change in AI capability, according to early reviewers. The model excels at tasks like editing, coding, and complex reasoning, often outperforming predecessors Opus 4.8 and rivals like GPT-5.5. However, the model is significantly slower and more expensive, and Anthropic warns it may no longer be included in standard subscriptions after June 22, pushing users toward per-token pricing. The system card also reveals that Anthropic intentionally restricts access to Fable for certain high-risk uses: anyone asking biological weapon–related questions, performing frontier ML research that rivals could exploit, or engaging in cybersecurity tasks will be silently downgraded to Opus 4.8. These classifiers fire about 5% of the time, even on benign queries, frustrating power users. The alternative—using Opus 4.8—offers lower cost and unrestricted access, but lacks Fable's advanced reasoning.

Beyond capability, the card outlines Anthropic's cautious approach to alignment and safety. Fable scores high on agentic autonomy tests (VendBench) and shows improved jailbreak robustness, but the company admits the safeguards are “not especially precise” and prioritize avoiding false negatives. Users must also accept a 30-day data retention policy for logging. While Fable is a breakthrough for professional use, the trade-offs in speed, price, and restrictions mean it’s not the right tool for every job. For many routine tasks, Opus 4.8 remains a more practical choice. Still, for high-stakes analysis or creative work, Fable's performance justifies the overhead.