Fable 5 Raises the Bar for AI Ambition

The episode covers the launch of Claude Fable 5, Anthropic's first Mythos-class model, which introduces a new naming tier above Opus. A restricted version called Claude Mythos 5 was also released exclusively to Project Glasswing partners in collaboration with the U.S. government. The host notes the unusually short interval between Opus 4.8 and Fable 5, and highlights how Anthropic's decision to use a full new number designation signals confidence in the leap forward.

Benchmark performance was highlighted as genuinely significant rather than marginal. Fable 5 dramatically outperformed GPT-5.5 across multiple benchmarks: Exploit Bench (78% vs 34%), Legal Agent Benchmark (13.3% vs 2.1%), SweeBench Pro (80.3% vs 58.6%), and the new Frontier Code benchmark (29.3% vs 5.7%). The Frontier Code benchmark, developed by Cognition, is designed to test not just whether code works but whether it is production-mergeable quality — addressing what was described as a major flaw in existing benchmarks where passing results often constitute 'unmergable slop.'

Three major controversies were discussed. First, Fable 5 automatically routes biology, chemistry, and cybersecurity questions to Opus 4.8 rather than answering them directly, with users reporting that even basic biology questions like 'what is mitochondria' or the word 'cancer' trigger fallbacks. The host acknowledged real frustration from actual biologists while dismissing what he characterized as performative outrage from people who only asked biology questions to confirm the problem. Second, buried in the system card, Anthropic revealed it has intentionally degraded the model's effectiveness for frontier LLM research tasks — including pre-training pipelines and distributed training infrastructure — to prevent competitors from using Claude to develop rival models. Researchers from Prime Intellect and others criticized this as anti-competitive and particularly objectionable because users are not informed when degradation occurs. Third, a mandatory 30-day data retention and human review policy for all Mythos-class model interactions raises serious enterprise concerns, particularly around NDAs when memory features are enabled.

On pricing, Fable 5 is set at $10 per million input tokens and $50 per million output tokens — double Opus pricing but less than half the cost of Mythos Preview through Project Glasswing. Access through subscription plans is described as a temporary introductory offer, with pay-per-usage required after June 23rd, reinforcing an industry-wide shift to usage-based pricing. Token efficiency concerns were raised by some users, but others argued that higher first-attempt success rates make the model cheaper in practice.

The host shared early use cases demonstrating transformative performance: Stripe reportedly compressed two months of engineering work on a 50-million-line Ruby codebase migration into a single day. Riley Brown one-shotted a functional Replit mobile clone in two prompts. The host himself used Fable 5 to rebuild Superintelligent's interview platform with a Whisper-based input model, create a new marketing and delivery platform for an enterprise AI training program, and transform a partially specced web experience for the AI Daily Brief into a production-ready pipeline.

The host concludes by arguing that Fable 5 signals a third era of AI use: moving from answers to tasks to responsibilities, where AI is given ongoing loops of work rather than discrete prompts. He introduces the concept of 'task imagination' — the emerging skill of identifying work ambitious enough to leverage models capable of running for hours or days — and suggests that most current AI training programs remain too focused on incremental improvements to existing workflows rather than genuinely rethinking what can be delegated.