Anthropic's Fable is the most capable AI model released to the public, and it comes with a hard ceiling. Stripe used it to migrate a 50-million-line Ruby codebase in a single day and completed a multi-thousand-line refactor in 45 minutes. In benchmark testing, Fable added 10 to 15 percentage points on key metrics, against an industry norm of 2 points. The performance is real.
The constraints are equally real. Guardrails trip on plant cell descriptions, LLM architecture questions, and software security topics. Tunguz frames this not as a flaw but as a structural necessity: critical infrastructure in banking, energy, and technology needs time to harden against the attacks that models this powerful will eventually enable. The ceiling was built on purpose.
The argument worth reading in full is not about what Fable can do. It is about the pacing problem. Techniques like RAG, MCP, and structured prompting are cycling in and out of fashion on a timeline measured in days. The ceiling will rise, Tunguz argues, but the question of how fast it should rise, and who decides, is where this piece gets genuinely uncomfortable.
[READ ORIGINAL →]