Summary
Matthew Berman delivers a detailed breakdown of Anthropic’s Fable 5.1 and Mythos 5.1 releases, working through pricing mechanics, benchmark results, enterprise data policies, and the model’s updated safety guardrails. While Anthropic frames the launch as a significant cost reduction, Berman notes that list prices remain $10 per million input tokens and $50 per million output tokens — unchanged from Fable 5. The stated 25% savings for typical workflows and up to 45% for highly agentic workloads come from reduced cache read pricing, not lower base rates.
On benchmarks, Fable 5.1 posts meaningful gains across TerminalBench Science, CursorBench, and computer use evaluations. A headline result: Fable 5.1 at low reasoning effort outperforms Fable 5 at maximum effort on several tasks while costing roughly a third as much — approximately $11 versus $34 per run on TerminalBench Science. The GDP-Val benchmark, created by the OpenAI team, shows a 130-point leap over Fable 5 and places Fable 5.1 above Opus 5 with a substantial gap over GPT-5.6 Soul. CursorBench shows a jump from 70.5% to 73.4% at max effort alongside a drop from $17.32 to $9.64 per run.
Berman also scrutinizes the new Enterprise Frontier Safeguards (EFS) and zero data retention feature, which store data in customer-controlled cloud infrastructure while Anthropic retains read access — a setup he calls a “half measure” for enterprises with strict data residency requirements. The episode closes on Anthropic’s controversial position that knowledge distillation from frontier models constitutes a safety risk, which Berman argues is functionally an argument against open-source AI development.
📺 Source: Matthew Berman · Published September 01, 2026
🏷️ Format: News Analysis
