Skip to content

Commit 40ba98e

Browse files
committed
bench(perf): the take-home's resumed leg, 1132 cycles
1 parent ae7b21a commit 40ba98e

1 file changed

Lines changed: 5 additions & 3 deletions

File tree

‎bench/perf/README.md‎

Lines changed: 5 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -124,10 +124,12 @@ standing goal** (`run_perf_chatgpt.sh`, official `score.sh`):
124124
|---|---|---|---|---|---|---|---|
125125
| 2026-08-20 | agent6 · gpt-5.6-sol max, standing goal | 1270 | 116.33× | 8 | 87 min | 215k / 116k | yes |
126126
| 2026-09-05 | agent6 · gpt-5.6-sol max, standing goal | 1215 | 121.59× | 8 | 58 min | 303k / 117k | yes |
127+
| 2026-09-05 | the same run resumed for a second leg | 1132 | 130.51× | 8 | +99 min | 3.76M / 94k | yes |
127128

128-
Plan-metered, $0 (the second run: 10 points of the plan, ended at the
129-
200-iteration cap); the best of every prior run here is 26.1×, and the
130-
upstream tests' eleven-hour Opus 4.5 mark is 1487.
129+
Plan-metered, $0 (the second run: 10 points of the plan for its first leg
130+
and 7 for the resumed one, each leg ending at the 200-iteration cap); the
131+
best of every prior run here is 26.1×, and the upstream tests' eleven-hour
132+
Opus 4.5 mark is 1487.
131133

132134
> **Update (per-call output cap, 32768 → 65536).** Three fresh kimi-k2.6
133135
> runs all landed at 1.0×, but with two distinct failure modes, both

0 commit comments

Comments
 (0)