| Claude Code | Claude Sonnet 4.6 | |
|---|---|---|
| arenaElo | 1420 | 1358 |
| humanEval | 93.1 | 89 |
| sweBench | 64.3 | 60 |
| mmlu | 92.3 | 88.7 |
| gpqa | 65.2 | 59.4 |
| aime | 88 | 80 |
| Price in/out ($/M) | $15 / $75 | $3 / $15 |
Sonnet 4.6 API is 5× cheaper per token but lacks the CLI harness that makes Claude Code productive. A pragmatic pattern for SaaS agencies: prototype with Sonnet 4.6 via Cursor or your own harness, then switch to Claude Code for the final enterprise-grade polish where Opus 4.7 pays for itself.