Cowerx

Back to daily.md

PirateBench

Agents build slices of our real projects from an empty workspace; a rubric grades them. Scores are out of 100, after intervention penalties. Every run of the two newest scored versions is listed, failures included; older and calibration epochs are summarized. 167 runs. Claude Code runs before v3 saw our notes file (no answers).

Build tiersscored v3

Tiers use the worst complete seed try: planning / inline 95+, quality worker 75+, budget worker 50+, otherwise not for building; correct = 95+, with cost and time across all complete tries per correct build at list prices.

modeltierworst try$/correct buildtime/correct buildcache share
gpt-6.1-solplanning / inline98.9$0.5725.9 min48%
opus-5-5planning / inline95.4$3.7129.8 min95%
gpt-6-solquality worker80.4none correct; $0.66/build–48%
gpt-6-lunanot for building45.9none correct; $0.02/build–47%
haiku-4-5not for building3.2none correct; $1.08/build–98%

Leaderboardscored v3

Out of 100 on the Kalshi data collection, T slices, items K01–K40, T01–T36; scored runs, rubric v3
epoch 68781f26be8f-de803160205e-0f6ff1c49115-7698d52d5555

modelrunsscoremedian runrange$/run
gpt-6.1-sol698.999.097.9–100.0$0.00
list $0.29
opus-5-5696.596.594.3–100.0$0.00
list $1.86
gpt-6-sol691.688.772.8–96.5$0.00
list $0.33
gpt-6-luna665.263.037.9–85.1$0.00
list $0.01
haiku-4-564.35.22.0–19.2$0.00
list $0.54

Score = each slice's median run, weighted by its items (K 40, T 36).

Score per slice

Kalshi data collection (K), median of runsgpt-6.1-sol97.9opus-5-596.4gpt-6-sol87.2gpt-6-luna74.7haiku-4-56.3T (T), median of runsgpt-6.1-sol100.0gpt-6-sol96.5opus-5-596.5gpt-6-luna54.7haiku-4-52.0

All runs (30)

2026 date (UTC)modelsliceseedscorepass/fail/part/totalnotes
10-02 20:10haiku-4-5T02.01/35/0/36
10-02 20:01haiku-4-5K04.22/38/0/40
09-30 16:21haiku-4-5T22.01/35/0/36
09-30 16:13haiku-4-5T119.27/28/1/36
09-30 15:53opus-5-5T296.535/1/0/36
09-30 15:45opus-5-5T196.535/1/0/36
09-30 15:37haiku-4-5K16.33/37/0/40
09-30 15:25gpt-6-solK290.236/4/0/40
09-30 15:14gpt-6.1-solT2100.036/0/0/36
09-30 15:06opus-5-5T096.535/1/0/36
09-30 14:58gpt-6-solK187.236/3/1/40
09-30 14:47gpt-6-lunaT254.120/16/0/36
09-30 14:34haiku-4-5K26.33/37/0/40
09-30 14:22gpt-6-solK077.432/7/1/40
09-30 14:12gpt-6-lunaT154.721/15/0/36
09-30 13:53gpt-6-lunaK285.135/5/0/40
09-30 13:42gpt-6-lunaT071.427/9/0/36
09-30 13:34opus-5-5K096.439/1/0/40
09-30 13:23gpt-6-lunaK137.916/24/0/40
09-30 13:12gpt-6.1-solT0100.036/0/0/36
09-30 13:04opus-5-5K194.338/2/0/40
09-30 12:53gpt-6-lunaK074.731/8/1/40
09-30 12:42gpt-6.1-solT1100.036/0/0/36
09-30 12:34gpt-6-solT296.535/1/0/36
09-30 12:26opus-5-5K2100.040/0/0/40
09-30 12:15gpt-6.1-solK297.939/1/0/40
09-30 12:04gpt-6-solT172.827/9/0/36
09-30 11:56gpt-6.1-solK197.939/1/0/40
09-30 11:45gpt-6-solT096.535/1/0/36
09-30 11:37gpt-6.1-solK097.939/1/0/40

Leaderboardscored v2

Out of 100 on the Kalshi data collection, MathForge server core, Recipe Kroger cart slices, items K01–K40, M01–M30, R01–R30; scored runs, rubric v2
epoch bc153d1e05b8-cfd29c413b22-7c295401355c-aba3dbce9d85

modelrunsscoremedian runrange$/run
opus-5-59 (9 regraded)91.290.278.0–100.0$0.00
list $2.35
gpt-6-sol9 (9 regraded)81.892.346.9–97.1$0.00
list $0.41
deepseek-v4-flash-07319 (9 regraded)62.890.24.2–95.1$0.11
mimo-v2.6-flash3 (3 unscored)–––$0.08
mimo-v2.6-pro6 (6 unscored)–––$0.18

Score = each slice's median run, weighted by its items (K 40, M 30, R 30).

Score per slice

Kalshi data collection (K), median of runsopus-5-581.5gpt-6-sol63.9deepseek-v4-flash-073118.2MathForge server core (M), median of runsopus-5-5100.0gpt-6-sol95.1deepseek-v4-flash-073194.8Recipe Kroger cart (R), median of runsopus-5-595.1gpt-6-sol92.3deepseek-v4-flash-073190.2
All runs (63)
2026 date (UTC)modelsliceseedscorepass/fail/part/totalnotes
09-28 15:45opus-5-5R2100.030/0/0/30
09-28 15:45opus-5-5R195.129/1/0/30
09-28 15:44opus-5-5R090.228/2/0/30
09-28 15:44gpt-6-solR297.129/1/0/30
09-28 15:44gpt-6-solR192.328/2/0/30
09-28 15:44gpt-6-solR089.427/3/0/30
09-28 15:43deepseek-v4-flash-0731R290.228/2/0/30
09-28 15:43deepseek-v4-flash-0731R188.526/4/0/30
09-28 15:43deepseek-v4-flash-0731R091.427/3/0/30
09-28 15:43opus-5-5M2100.030/0/0/30
09-28 15:43opus-5-5M189.427/3/0/30
09-28 15:42opus-5-5M0100.030/0/0/30
09-28 15:42gpt-6-solM295.129/1/0/30
09-28 15:42gpt-6-solM195.129/1/0/30
09-28 15:41gpt-6-solM095.129/1/0/30
09-28 15:41deepseek-v4-flash-0731M295.129/1/0/30
09-28 15:41deepseek-v4-flash-0731M194.628/1/1/30
09-28 15:41deepseek-v4-flash-0731M094.828/1/1/30
09-28 15:40opus-5-5K281.534/6/0/40
09-28 15:36opus-5-5K181.534/6/0/40
09-28 15:30opus-5-5K078.033/7/0/40
09-28 15:26gpt-6-solK246.921/19/0/40
09-28 15:21gpt-6-solK183.033/6/1/40
09-28 15:17gpt-6-solK063.927/13/0/40
09-28 15:12deepseek-v4-flash-0731K24.22/38/0/40
09-28 15:07deepseek-v4-flash-0731K118.28/32/0/40
09-28 15:03deepseek-v4-flash-0731K020.39/31/0/40
09-28 10:20mimo-v2.6-proK2–28/11/1/40harness error: wrong route (OpenRouter, not Xiaomi direct)
09-28 09:17mimo-v2.6-proK1–30/9/1/40timeout; harness error: wrong route (OpenRouter, not Xiaomi direct)
09-28 08:12mimo-v2.6-proK0–31/8/1/40timeout; harness error: wrong route (OpenRouter, not Xiaomi direct)
09-28 06:45mimo-v2.6-proK1–0/40/0/40harness error: provider rate limit: Provider rate limit reached. Please retry shortly. (request_id=req_28544fd59f16)
09-28 06:41mimo-v2.6-proK0–0/40/0/40harness error: provider rate limit: Provider rate limit reached. Please retry shortly. (request_id=req_86de6842cc8d)
09-28 06:39mimo-v2.6-proK2–0/40/0/40harness error: provider rate limit: Provider rate limit reached. Please retry shortly. (request_id=req_dc3769037d1e)
09-28 03:30mimo-v2.6-flashK0–0/40/0/40timeout; harness error: wrong route (OpenRouter, not Xiaomi direct)
09-28 03:30mimo-v2.6-flashK1–0/40/0/40timeout; harness error: wrong route (OpenRouter, not Xiaomi direct)
09-28 03:30mimo-v2.6-flashK2–0/40/0/40timeout; harness error: wrong route (OpenRouter, not Xiaomi direct)
09-27 17:50opus-5-5R2–30/0/0/30not counted: busy machine (load1 30.2)
09-27 17:50opus-5-5R1–29/1/0/30not counted: busy machine (load1 27.3)
09-27 17:49opus-5-5R0–28/2/0/30not counted: busy machine (load1 28.0)
09-27 17:49gpt-6-solR2–29/1/0/30not counted: busy machine (load1 25.3)
09-27 17:49gpt-6-solR1–28/2/0/30not counted: busy machine (load1 21.4)
09-27 17:48gpt-6-solR0–27/3/0/30not counted: busy machine (load1 21.2)
09-27 17:48deepseek-v4-flash-0731R2–28/2/0/30not counted: busy machine (load1 18.8)
09-27 17:48deepseek-v4-flash-0731R1–26/4/0/30not counted: busy machine (load1 18.2)
09-27 17:47deepseek-v4-flash-0731R0–27/3/0/30not counted: busy machine (load1 18.9)
09-27 17:47opus-5-5K2–34/6/0/40not counted: busy machine (load1 28.4)
09-27 17:42opus-5-5K1–34/6/0/40not counted: busy machine (load1 24.5)
09-27 17:36opus-5-5K0–30/9/1/40not counted: busy machine (load1 38.7)
09-27 17:32gpt-6-solK2–20/20/0/40not counted: busy machine (load1 80.2)
09-27 17:27gpt-6-solK1–26/14/0/40not counted: busy machine (load1 58.0)
09-27 17:22gpt-6-solK0–27/13/0/40not counted: busy machine (load1 19.1)
09-27 17:17gpt-6-solM1–29/1/0/30not counted: busy machine (load1 22.0)
09-27 17:16gpt-6-solM2–29/1/0/30not counted: busy machine (load1 20.0)
09-27 17:16deepseek-v4-flash-0731K2–2/38/0/40not counted: busy machine (load1 25.4)
09-27 17:11gpt-6-solM0–29/1/0/30not counted: busy machine (load1 26.6)
09-27 17:10deepseek-v4-flash-0731K1–8/32/0/40not counted: busy machine (load1 28.6)
09-27 17:05deepseek-v4-flash-0731M2–29/1/0/30not counted: busy machine (load1 27.3)
09-27 17:05opus-5-5M2–30/0/0/30not counted: busy machine (load1 25.0)
09-27 17:04opus-5-5M1–27/3/0/30not counted: busy machine (load1 23.6)
09-27 17:03deepseek-v4-flash-0731K0–9/31/0/40not counted: busy machine (load1 29.5)
09-27 16:51opus-5-5M0–30/0/0/30not counted: busy machine (load1 30.7)
09-27 16:46deepseek-v4-flash-0731M0–28/1/1/30not counted: busy machine (load1 34.1)
09-27 16:40deepseek-v4-flash-0731M1–28/1/1/30not counted: busy machine (load1 41.7)
Older and calibration epochs (5)
  • scored v1 c4dbfc7e5c5a-644f71243d5d-c182c64574b7-88e7f43dcd6d · opus-5-5 49.9 · 24 runs
  • calibration c4dbfc7e5c5a-644f71243d5d-c182c64574b7-88e7f43dcd6d · opus-5-5 · harness 2 75.0 · 20 runs
  • calibration 46fc803f1a9c-ec54f73daa8b-c182c64574b7 · deepseek-v4.1-flash (run as deepseek-v4-flash-0731) 60.0 · 10 runs
  • calibration fe823afec90b-7c33790e499b-c182c64574b7 · deepseek-v4.1-flash (run as deepseek-v4-flash-0731) 65.0 · 9 runs
  • unknown unknown · deepseek-v4-flash-0731 – · 11 runs