Anthropic's 'Too Dangerous' AI is Here: Fable 5 Unleashed

Summary

Anthropic has released Fable 5, a new frontier AI model previously deemed too dangerous for public release, now made safe with guardrails. This model represents a new class of AI, exceeding the capabilities of previous generally available models. Benchmarks show Fable 5 significantly outperforming Opus 4.8 and GPT 5.5 in areas like coding with an 80% SWEBench Pro score, compared to 69% and 58% respectively. For coding tasks, it achieved 29.3% on Frontier Code Diamond, vastly exceeding Opus 4.8 and GPT 5.5. Fable 5 also excels in real-world knowledge work, scoring 1932 on GDP valve, higher than GPT 5.5's 1769. The model demonstrates advanced tool use and computer interaction, with a 13% score on the Legal Agent Benchmark, up from 10% for Opus. Fable 5 is a 10 trillion parameter model, showing exceptional performance in software engineering, knowledge work, and scientific research, particularly on longer and more complex tasks. It compressed months of engineering work into days on a 50 million line Ruby codebase, performing a migration in one day that would normally take a team over two months. Pricing is set at $10 per million input tokens and $50 per million output tokens, a high cost justified by its advanced capabilities for the most difficult problems. Fable 5 is more token efficient, but its output is extremely verbose and information-dense, sometimes making it difficult to read and requiring simplification prompts. It also exhibits a tendency to ask numerous clarifying questions before executing tasks, which can be frustrating. Despite these quirks, Fable 5 demonstrates impressive autonomy and the ability to leverage workflows and parallel agent execution effectively. The model can even perform tasks like completing the game Pokemon Fire Red using only vision, and creating complex simulations like fluid dynamics in real-time. Anthropic has implemented a 30-day data retention policy for Mythos-class models to enhance safety and defend against attacks.