Prove the lift · AI model performance evaluation

claude-sonnet-4-5 vs claude-3-7-sonnet

-10-point lift for claude-sonnet-4-5 on code-review-bench

Test summary

This benchmark evaluates claude-sonnet-4-5 against claude-3-7-sonnet on the code-review-bench v3 dataset (7 rows), scored by an LLM-as-judge (gpt-4o-mini). claude-sonnet-4-5 reached a 21% pass rate versus 31% for claude-3-7-sonnet — a -10-point lift.

Base model
claude-3-7-sonnet
via anthropic · 31% pass rate
Refined model
claude-sonnet-4-5
via anthropic · 21% pass rate · 3.0s p95
Base pass rate
31%
Refined pass rate
21%
Refined p95
3.0s
Est. tokens saved
—
−0pts
no lift — the base model held its ground
0%
Base model
0%
Invoked-refined
▲2 improved▼0 regressed●5 unchanged

Per-row results

1Spot a race in a read-modify-write on shared state0% → 0%0
2Identify an N+1 query in an ORM call site0% → 100%+100
3Note a broken null-check after a refactor0% → 0%0
4Catch an unsanitized input reaching a SQL string0% → 100%+100
5Find a resource leak on an early-return path0% → 0%0
6Detect a missing await on an async DB write0% → 0%0
7Flag an off-by-one in a pagination loop100% → 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