Claude Sonnet 5: Benchmarks and Pricing vs Opus 4.8

My notes

Summary

Quick benchmark-and-pricing rundown of Claude Sonnet 5, the first Sonnet upgrade since 4.6. It beats Opus 4.8 on knowledge work but trails on coding, positioned as a cheaper workhorse for simpler tasks rather than a full Opus replacement.

Key Insight

• Pricing: $2 / M input tokens, $10 / M output tokens, roughly 60% cheaper than Opus.
• Where it wins: knowledge work, where Sonnet 5 actually beats Opus 4.8.
• Where it loses: coding, around 2% behind on Terminal Bench 2.1 and around 6% behind on agentic coding versus Opus 4.8.
• Counterintuitive cost point: on complex tasks Opus can be cheaper despite the higher per-token price, because it is more token-efficient (fewer tokens to reach the answer). Sticker price per token is not the real cost.
• Framed as a “great workhorse,” reserving Opus (or Fable) for the tasks that genuinely need it.