The two models share a 1,000,000-token context window, and Opus 4.8 leads on every benchmark: Intelligence Index 55.7 vs 53.4, Coding Index 74.3 vs 71.5, and Agentic Index 47.2 vs 46.7. The gaps are real but not large enough to justify the price difference on every step. Opus 4.8 costs $5/$25 per million tokens; Sonnet 5 costs $2/$10. Paying 2.5x more on output earns its keep only on the specific steps where that extra quality differential shows.
Claude Sonnet 5 launched June 30, 2026 and is the newer model. Anthropic built in selectable reasoning effort (low, medium, high, max), so you can tune how hard Sonnet 5 thinks on each step. A classification call at low effort is cheap and fast; a complex analysis call at max effort approaches Opus territory on tasks where reasoning depth is the variable. The adaptive thinking feature is the mechanism that partially closes the benchmark gap without paying Opus prices.
Put Sonnet 5 on every step where output is verified programmatically or where draft quality is sufficient. Put Opus 4.8 only on the steps where the quality differential is noticeable to the downstream consumer: final reasoning, contract-level analysis, complex code generation that does not run through tests, and decisions a human reads without a check. Most agent workflows have two or three Opus-class steps and many more Sonnet-class steps. Identify the Opus steps first, then route everything else to Sonnet 5.