Opus 5 vs Fable 5: claimed benchmark wins at half the price

My notes

Summary

A short creator take on the Opus 5 release, claiming it beats Fable 5 on most benchmarks except health and legal domains while costing roughly half as much per token. The video also claims Opus 5 applies fewer safety guardrails on sensitive topics like cybersecurity and biology than the comparison model.

Key Insight

• Claimed pricing: $5 per million input tokens and $25 per million output tokens, pitched as roughly 50% cheaper than the rival model at similar quality.
• Claimed benchmark pattern: the rival model still wins on health and legal-domain tasks, and even on multidisciplinary reasoning once tool calls are involved, but loses on most general benchmarks.
• Efficiency framing: unlike a prior case where a “cheaper” model turned out token-inefficient in practice, raising real-world cost despite the lower sticker price, this model is claimed to be efficient on both list price and actual token usage.
• Guardrail claim: fewer refusals or downgrades to a weaker model on cybersecurity and biology questions, but explicitly not as strong as top models at finding zero-day exploits, which the creator frames as the reason looser guardrails are tolerated.
• No official benchmark source, methodology, or named benchmark suite is cited in the video, so treat all figures as a single creator’s unverified claims, not primary-source data.