TLDRocket
23 July 2026
Anthropic's release of Claude Fable 5 exposed a fundamental tension in AI development: the gap between what models can do and what companies allow them to do. Fable's safety filters rerouted 8 to 35 percent of flagged tasks to weaker models, inflating benchmark scores when fallback mechanisms were counted but collapsing performance when refusals were marked as failures. Independent evaluators found it impossible to measure true capability, and U.S. export controls further restricted assessment. The friction has accelerated demand for open alternatives. Poolside AI released Laguna S 2.1, a 118-billion-parameter model that outperforms Deepseek's larger competitors on coding benchmarks while requiring less compute—proof that engineering rigor can compete with scale. Meanwhile, Mistral abandoned the frontier model race entirely, pivoting to enterprise services where revenue grew 37 percent year-over-year, with 80 percent now from deployment and customization rather than model licensing.
Read the full briefing →