Prove the lift · AI model performance evaluation

gpt-4o vs claude-sonnet-4-6

+2-point lift for gpt-4o on code-review-bench

Test summary

This benchmark evaluates gpt-4o against claude-sonnet-4-6 on the code-review-bench v3 dataset (7 rows), scored by an LLM-as-judge (gpt-4o-mini, run locally). gpt-4o reached a 50% pass rate versus 48% for claude-sonnet-4-6 — a +2-point lift.

Base model
claude-sonnet-4-6
via anthropic · 48% pass rate
Refined model
gpt-4o
via openai · 50% pass rate · 2.2s p95
Base pass rate
48%
Refined pass rate
50%
Refined p95
2.2s
Est. tokens saved
6.9K
+0pts
refined beats base on 7 rows
0%
Base model
0%
Invoked-refined
▲3 improved▼2 regressed●2 unchanged

Per-row results

1Flag an off-by-one in a pagination loop0% → 100%+100
2Find a resource leak on an early-return path100% → 0%-100
3Spot a race in a read-modify-write on shared state100% → 0%-100
4Detect a missing await on an async DB write100% → 100%0
5Catch an unsanitized input reaching a SQL string0% → 100%+100
6Note a broken null-check after a refactor0% → 100%+100
7Identify an N+1 query in an ORM call site100% → 100%0
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