claude-code + claude-haiku-4-5 vs claude-code + claude-opus-5
Each agent graded on the same private suites, running headless through its own tools. Deltas are oriented so a win is a win, whichever way the metric runs.
| Dimension | claude-code + claude-haiku-4-5 | claude-code + claude-opus-5 | Delta |
|---|---|---|---|
| Capability | 30% | 60% | 30% · claude-code + claude-opus-5 |
| Median latency | 188763 ms | 222605 ms | 33842 ms · claude-code + claude-haiku-4-5 |
| Reliability | 100% | 100% | even |
| Error rate | 0% | 0% | even |
claude-code + claude-haiku-4-5: 17.0% of Claude Max 5x per run, about 5.9 runs per 5h. claude-code + claude-opus-5: 64.0% of Claude Max 5x per run, about 1.6 runs per 5h. Each figure is a share of that agent's own plan, so there is no delta to draw.
Work per window: claude-code + claude-haiku-4-5 859 vs claude-code + claude-opus-5 228 suite runs per month (claude-code + claude-haiku-4-5 leads); per plan-dollar: 8.59 vs 2.28 (claude-code + claude-haiku-4-5 leads). Work figures compare outputs, so unlike quota shares they are comparable across plans.
Grades and deltas are comparable only within a suite version. See methodology.