GPT 6 Luna max leads on quality and cost: 2.2 points higher and 1.3× cheaper; DeepSeek V4.1 Flash leads on agent time: 1.2× faster.
Stet score
DeepSeek V4.1 FlashRank 9 of 14
47.895% range 37.5–58.0
GPT 6 Luna maxRank 8 of 14
50.095% range 40.1–59.32.2 points higher
DeepSeek V4.1 Flash solved 1 task that GPT 6 Luna max didn’t, and GPT 6 Luna max solved 1 that DeepSeek V4.1 Flash didn’t. Both solved 11; neither solved 11.
stet-15439c21Passed testsstet-commit-f8ae1177Passed testsA task counts as solved when its tests passed or a judge found the fix equivalent to the merged change, as in the leaderboard’s Solved column. Stet score also weighs code quality, so it can order the two differently.
| Task | DeepSeek V4.1 Flash | GPT 6 Luna max | Solved by |
|---|---|---|---|
stet-3169aefa | Passed tests | Passed tests | Both |
stet-89dfbc27 | Passed tests | Passed tests | Both |
stet-b28f4fe7 | Passed tests | Passed tests | Both |
stet-232b50ba | Passed tests | Passed tests | Both |
stet-5276aa3f | Passed tests | Passed tests | Both |
stet-commit-bbbbae09 | Passed tests | Passed tests | Both |
stet-15439c21 | Passed tests | Failed | DeepSeek V4.1 Flash only |
stet-66762b4c | Passed tests | Equivalent to the merged fix | Both |
stet-commit-f0caded9 | Passed tests | Equivalent to the merged fix | Both |
stet-commit-5d766dde | Passed tests | Equivalent to the merged fix | Both |
stet-43493c52 | Passed tests | Passed tests | Both |
stet-6f84e978 | Failed | Failed | Neither |
stet-commit-f8ae1177 | Failed | Passed tests | GPT 6 Luna max only |
stet-2450ca2d | Failed | Failed | Neither |
stet-commit-342d0ac0 | Equivalent to the merged fix | Equivalent to the merged fix | Both |
stet-commit-60a0e6c2 | Failed | Failed | Neither |
stet-commit-5e075641 | Failed | Failed | Neither |
stet-commit-2826765e | Failed | Failed | Neither |
stet-commit-a8716b88 | Failed | Failed | Neither |
stet-commit-30506f8e | Failed | Failed | Neither |
stet-commit-87814583 | Failed | Failed | Neither |
stet-commit-d2ab0c25 | Failed | Failed | Neither |
stet-commit-e41c2e2d | Failed | Failed | Neither |
stet-commit-e928166f | Failed | Failed | Neither |
| Metric | DeepSeek V4.1 Flash | GPT 6 Luna max |
|---|---|---|
| Rank | 9 of 14 | 8 of 14 |
| Stet score | 47.895% range 37.5–58.0 | 50.095% range 40.1–59.3 |
| Solved | 12/2411 passed, 1 judged equivalent | 12/248 passed, 4 judged equivalent |
| Passed review | 8/24 | 6/24 |
| Matched merged change | 10/231 not recorded | 11/24 |
| Cost / task | $0.49from 23/24 tasks | $0.37 |
| How cost was recorded | estimated from tokens on 23, not recorded on 1 | reported by its harness |
| Agent time / task | 49.3 min | 59.7 min |
| Ran in | Daytona sandbox | Daytona sandbox |
The leaderboard’s readings on groups of tasks. Each group holds fewer tasks than the overall standings, so treat these as leads to test on your repo.
11 tasks
11 tasks whose merged change was 2–149 lines in 1–2 files. The 7 small fixes below are among them.
Both matched the merged change on 7 of 11.
DeepSeek V4.1 Flash passed review more often: 8 of 11 to GPT 6 Luna max’s 6.
| DeepSeek V4.1 Flash | GPT 6 Luna max | |
|---|---|---|
| Matched merged change | 7/11 | 7/11 |
| Passed review | 8/11 | 6/11 |
| Cost / task | $0.21 | $0.09 |
| Agent time / task | 34.9 min | 19.0 min |
7 tasks
7 tasks whose instructions start with “Fix”. Their merged changes were 4–29 lines in 1–2 files.
Both matched the merged change on 5 of 7.
DeepSeek V4.1 Flash passed review more often: 6 of 7 to GPT 6 Luna max’s 4.
| DeepSeek V4.1 Flash | GPT 6 Luna max | |
|---|---|---|
| Matched merged change | 5/7 | 5/7 |
| Passed review | 6/7 | 4/7 |
| Cost / task | $0.19 | $0.08 |
| Agent time / task | 34.6 min | 17.1 min |
13 tasks
13 tasks whose merged change ran over 150 lines or more than 3 files: 113–1,420 lines in 2–19 files.
| DeepSeek V4.1 Flash | GPT 6 Luna max | |
|---|---|---|
| Matched merged change | 3/121 not recorded | 4/13 |
| Passed review | 0/13 | 0/13 |
| Cost / task | $0.751 not recorded | $0.61 |
| Agent time / task | 61.5 min | 94.2 min |
Stet replays your merged PRs, scores quality above pass/fail, and delivers comparison reports through your agent.
Run it on your repocompare-n24-r3b.