Descriptions:
Matthew Berman shares firsthand observations from early access to Claude Fable 5, which Anthropic describes as a Mythos-class model made safe for general release. The benchmark section is thorough: SWE-bench Pro scores 80% versus Opus 4.8 at 69% and GPT 5.5 at 58%; Frontier Code Diamond reaches 29.3% compared to roughly 14% for Opus and 5.7% for GPT 5.5; GDPVal hits 1932 against 1769 and 1890 for the competing models; Terminal Bench scores 83.4% versus 88% for Opus. Berman also notes this is reportedly the first 10-trillion-parameter model of its kind.
Beyond benchmarks, Berman describes the subjective experience of using Fable: the model tends to deeply explore entire codebases before responding, producing information-dense outputs that require slower reading than previous Claude versions. He introduces Ultra Code — Fable’s parallel workflow feature — which during one live test spun up over 100 simultaneous sub-agents to review a full codebase, producing a comprehensive report while visibly accelerating token consumption. His practical recommendation is to start at the lowest available thinking-effort setting, as even that level frequently exceeds what most tasks require.
The video closes with a speculative but interesting observation: if future AI models prioritized inter-model communication efficiency over human readability, they might develop hyperinformation-dense non-alphanumeric representations — raising interpretability and safety questions that Berman frames as a logical extension of what Fable’s output density already hints at.
📺 Source: Matthew Berman · Published June 09, 2026
🏷️ Format: Review
