Claude Fable 5
Points and comments are a snapshot, not live.
Anthropic releases Claude Fable 5, a state-of-the-art model with safeguards, and Claude Mythos 5 for authorized cybersecurity professionals.
Anthropic launched Claude Fable 5, described as state-of-the-art on most benchmarks in software engineering, knowledge work, vision, and scientific research. The model exceeds prior Claude versions on benchmarks including SWE-Bench (95.5%), GPQA Diamond (94.1%), and Humanity's Last Exam (59%). Fable 5 comes with safety classifiers that redirect queries on cybersecurity, biology, chemistry, and model development to Claude Opus 4.8, triggering in less than 5% of sessions. A parallel model, Claude Mythos 5, removes these safeguards for authorized government and infrastructure users via Project Glasswing. Both models cost $10 per million input tokens and $50 per million output tokens. Early testers reported Fable 5 compressed months of engineering work into days; Stripe cited a two-month codebase migration completed in one day. Mythos 5 showed 10x acceleration in protein design tasks and produced molecular biology hypotheses that scientists preferred over Opus-class alternatives in 80% of blind comparisons.
What commenters are saying
Commenters raised concerns about the release contradicting prior safety claims. One noted Anthropic previously stated Mythos was too dangerous for general availability, now released with classifier "safeguards" that critics describe as prompt-injectable. A key concern emerged around safeguards blocking frontier LLM development work (impacting ~0.03% of traffic), interpreted as Anthropic restricting use of its own model for competitor development. Comments highlighted that safety classifiers may be bypassable and questioned whether authorized Mythos access truly prevents misuse by bad actors. Technical details surfaced: benchmarks show small deltas between Fable and Mythos performance (e.g., SWE-Bench Pro: 80.3% vs 80.0%), and METR reported Mythos 5 cannot reliably automate R&D across multiple weeks. One commenter flagged that Anthropic's vetting of partner communications may shape public perception.