Is Claude Sonnet 5 as good as Opus?

Close — closer than the price gap suggests, which is exactly why the question matters. On the SWE-bench Verified coding benchmark (llm-stats, July 2026), Claude Sonnet 5 scores 85.2% against Claude Opus 4.8's 88.6% — a 3.4-point gap. Opus costs $5 per million input tokens and $25 per million output, versus Sonnet 5's introductory $2/$10 (standard $3/$15 from September 1, 2026).

At intro pricing, that's 2.5× the cost for a few points of benchmark performance.

Whether those points matter depends entirely on the work. For everyday production tasks — summarization, extraction, routine coding, agent steps with clear instructions — most teams find Sonnet-class models indistinguishable in practice, and the savings compound at volume. For the hardest work — long multi-step agentic coding, high-stakes reasoning, subtle analysis where a wrong answer is expensive — the flagship's edge shows up in exactly the cases you care most about.

And above both sits Claude Fable 5 ($10/$50, 95.0% on the same board) for the genuinely frontier tasks.

The pattern that's become standard as of 2026: route to Sonnet 5 by default, escalate to Opus or Fable on failure or on high-stakes paths. That gets you flagship quality where it counts and mid-tier economics everywhere else.