Sonnet 5 vs Opus 4.8 tested across 24 coding tasks at every reasoning level
Using 24 tasks drawn from actually-merged in two repos (graphql-go-tools in Go, sqlparser-rs in Rust), both Sonnet 5 and Opus 4.8 were run once each at five s — low, medium, high, xhigh, max — inside Claude Code, with a GPT-5.4 judge scoring the resulting code patches. As increased, Sonnet did more checking and produced longer working trajectories, and the judge rated its patches as clearer and more intentional. Opus's behavior stayed relatively flat across , and the judge favored its patches as simpler, more robust, and more minimal.
Neither approach is better across the board — they fail differently and work differently. Cost varied sharply with : at low effort Sonnet cost 0.62x Opus, at medium 0.81x, at high and xhigh the two were roughly tied, and at max Sonnet actually cost 1.37x more than Opus.
Key points
- 24 tasks pulled from real merged PRs across two repos: graphql-go-tools (Go) and sqlparser-rs (Rust)
- Sonnet 5 and Opus 4.8 each run once per (low, medium, high, xhigh, max), scored by a GPT-5.4 judge
- Sonnet responds to higher with more checking and longer trajectories; Opus stays comparatively flat
- Pricing: Sonnet costs 0.62x Opus at low, 0.81x at medium, roughly tied at high/xhigh, but 1.37x more expensive at max
- Judge favored Sonnet's patches as clearer/more intentional and Opus's as simpler/more robust and minimal