Summary
IndyDevDan covers two intertwined stories in this video: the sudden federal suspension of Claude Fable 5 and Mythos 5, and a detailed hands-on benchmark comparing those models against Opus and Sonnet across 15 agentic application builds.
On the regulatory side, the video reports that a federal export control order forced Anthropic to pull both Fable 5 and Mythos 5 from Pro, Team, and enterprise subscription tiers after the government identified a jailbreak. A notable detail: Anthropic has stated the same jailbreaking techniques work on GPT-5.5, making the selective suspension of Claude models — and not OpenAI’s — a point of active controversy. The creator characterizes the ban as temporary, with Fable expected to return.
The benchmark section is the core of the video. IndyDevDan ran 15 independent agent sandboxes — five each on Sonnet, Opus, and Fable 5 — each executing the same five application specs including a Simon Willison LLM price index clone, a Hacker News clone, a scikit-learn model generator, and an agent-builds-agent application. Results: Sonnet cost $55 using ~700K tokens, Opus cost $91 using ~700K tokens, and Fable 5 cost $200 using ~1 million tokens. Despite using more tokens and costing more, Fable 5 completed the same work approximately 20% faster. The creator argues the right metric for evaluating models like Fable is “price per intelligent agent hour” rather than cost per token, and positions Fable as an orchestrator-class model — best used to delegate to sub-agents on large, complex, long-running specs rather than as a drop-in replacement for Opus on routine tasks.
📺 Source: IndyDevDan · Published June 15, 2026
🏷️ Format: Benchmark Test
