Prove the lift · AI model performance evaluation
deepseek-r1 vs gpt-5
+17-point lift for deepseek-r1 on code-review-bench
Test summary
This benchmark evaluates deepseek-r1 against gpt-5 on the code-review-bench v2 dataset (7 rows), scored by an LLM-as-judge (llama-3.1-8b, run locally). deepseek-r1 reached a 62% pass rate versus 45% for gpt-5 — a +17-point lift.
- Base model
- gpt-5
- via openai · 45% pass rate
- Refined model
- deepseek-r1
- via deepseek · 62% pass rate · 2.0s p95
Base pass rate
45%
Refined pass rate
62%
Refined p95
2.0s
Est. tokens saved
4.0K
+0pts
refined beats base on 7 rows
0%
Base model0%
Invoked-refined▲0 improved▼0 regressed●7 unchanged
Per-row results
1Note a broken null-check after a refactor100% → 100%0
2Flag an off-by-one in a pagination loop0% → 0%0
3Catch an unsanitized input reaching a SQL string100% → 100%0
4Spot a race in a read-modify-write on shared state100% → 100%0
5Detect a missing await on an async DB write100% → 100%0
6Find a resource leak on an early-return path100% → 100%0
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 InvokedGenerated with Invoked · Browse all benchmarks