Prove the lift · AI model performance evaluation

claude-opus-4-6 vs claude-opus-4-8

+11-point lift for claude-opus-4-6 on code-review-bench

Test summary

This benchmark evaluates claude-opus-4-6 against claude-opus-4-8 on the code-review-bench v3 dataset (5 rows), scored by an LLM-as-judge (claude-haiku-4-5, run locally). claude-opus-4-6 reached a 51% pass rate versus 40% for claude-opus-4-8 — a +11-point lift.

Base model
claude-opus-4-8
via anthropic · 40% pass rate
Refined model
claude-opus-4-6
via anthropic · 51% pass rate · 0.8s p95
Base pass rate
40%
Refined pass rate
51%
Refined p95
0.8s
Est. tokens saved
11.1K
+0pts
refined beats base on 5 rows
0%
Base model
0%
Invoked-refined
▲4 improved▼0 regressed●1 unchanged

Per-row results

1Identify an N+1 query in an ORM call site0% → 100%+100
2Note a broken null-check after a refactor0% → 100%+100
3Flag an off-by-one in a pagination loop0% → 100%+100
4Catch an unsanitized input reaching a SQL string0% → 0%0
5Find a resource leak on an early-return path0% → 100%+100
Prove the lift on your own data.

Run any two models over your tasks, judged automatically, and see the improvement — in seconds.

Try Invoked
Generated with Invoked · Browse all benchmarks