BixBench Verified-50 · supplementary experiment · October 2026

Third-round token optimization

Galaxy token use is now 1.53× Open-ended code

On the same 50 BixBench tasks with GPT-5.6 Sol, three new Galaxy-API code replicates averaged 54.35M tokens, compared with 35.51M for Open-ended code. Previous website Galaxy runs averaged 122.41M, or 3.45× Open-ended code. The new mean is 55.6% lower than that previous Galaxy mean.

This is a separate supplementary batch. All earlier results and run records remain unchanged.

All three replicates

Replicate 1 · 50 tasks46/50 correct

52.50M tokens · 1.48× Open-ended code

Replicate 2 · 50 tasks46/50 correct

60.45M tokens · 1.70× Open-ended code

Replicate 3 · 50 tasks45/50 correct

50.09M tokens · 1.41× Open-ended code

Overall: 137/150 correct (91.3%). Tokens are input plus output; cached input is already included in input tokens.

Comparison with Open-ended code

Open-ended code35.51M
Previous Galaxy122.41M
Final Galaxy54.35M
ConfigurationTokens per complete 50-task runRelative to Open-endedCoverage
Open-ended code35.51M1.00×3-replicate mean
Previous website Galaxy122.41M3.45×3-replicate mean
Final optimized Galaxy54.35M1.53×3-replicate mean

All three rows average three replicates of the same 50 tasks with GPT-5.6 Sol. The Open-ended and previous Galaxy baselines reuse the existing website records. Previous runs used Codex CLI 0.146.0; the new Galaxy runs used 0.156.0.

What changed in the third optimization round

Full execution evidence is retained despite the shorter conversation.

Run configuration

Workers used only the benchmark's Galaxy tools; additional account-connected tools were disabled.

These are observed batch comparisons; savings were not measured separately for each individual change. Run selection is recorded in the batch manifest, and original attempts remain retained locally.

All 50 tasks and 150 run records

Open each item for submitted answers, tokens, HF traces and Galaxy histories. A Galaxy history can be opened by link but is not listed publicly.

ItemR1R2R3Open-ended meanNew Galaxy / Open-ended
bix-11-q1Correct
0.58M
Correct
0.77M
Correct
0.63M
0.28M2.36×
bix-11-q2Correct
0.72M
Correct
0.51M
Correct
0.57M
0.40M1.52×
bix-12-q2Correct
1.13M
Correct
2.18M
Correct
1.69M
0.44M3.76×
bix-12-q4Correct
2.96M
Correct
4.54M
Correct
2.20M
0.61M5.32×
bix-12-q5Correct
2.85M
Correct
2.39M
Correct
2.55M
0.50M5.15×
bix-12-q6Correct
1.60M
Correct
0.90M
Correct
2.07M
0.45M3.37×
bix-14-q1Correct
2.67M
Correct
2.61M
Incorrect
1.92M
0.50M4.81×
bix-16-q1Correct
2.06M
Correct
0.40M
Correct
0.48M
0.39M2.52×
bix-16-q3Correct
0.46M
Correct
0.41M
Correct
0.45M
0.30M1.47×
bix-16-q4Correct
0.62M
Correct
0.35M
Correct
0.56M
0.28M1.79×
bix-17-q2Correct
1.33M
Correct
1.15M
Correct
1.78M
0.36M4.00×
bix-18-q1Correct
0.36M
Correct
0.56M
Correct
0.61M
0.07M7.65×
bix-18-q3Correct
0.49M
Correct
0.32M
Correct
0.41M
0.08M4.97×
bix-20-q3Correct
1.19M
Correct
1.87M
Correct
1.67M
0.39M4.06×
bix-22-q1Correct
0.65M
Correct
0.60M
Correct
0.65M
0.19M3.29×
bix-22-q4Correct
0.90M
Correct
1.79M
Correct
0.83M
0.18M6.45×
bix-24-q2Correct
1.46M
Correct
3.31M
Correct
1.30M
1.52M1.33×
bix-26-q3Correct
0.72M
Correct
0.78M
Correct
1.48M
0.26M3.87×
bix-26-q5Correct
0.75M
Correct
0.82M
Correct
0.75M
0.43M1.80×
bix-27-q5Correct
0.71M
Correct
0.79M
Correct
0.92M
0.23M3.56×
bix-28-q3Correct
0.64M
Incorrect
0.62M
Correct
0.72M
0.52M1.27×
bix-30-q3Correct
2.22M
Correct
1.31M
Correct
0.81M
0.56M2.59×
bix-31-q2Correct
0.80M
Correct
0.59M
Correct
0.70M
12.70M0.05×
bix-32-q2Incorrect
1.31M
Correct
1.38M
Incorrect
1.42M
0.48M2.83×
bix-34-q2Correct
0.93M
Correct
1.85M
Correct
0.82M
0.21M5.64×
bix-34-q5Correct
0.70M
Correct
0.69M
Correct
0.87M
0.24M3.10×
bix-35-q1Correct
0.61M
Correct
0.69M
Correct
0.80M
0.32M2.19×
bix-35-q2Correct
0.58M
Correct
0.63M
Correct
0.57M
0.59M1.00×
bix-37-q1Correct
0.62M
Correct
0.52M
Correct
0.56M
0.13M4.35×
bix-37-q4Correct
0.44M
Correct
0.56M
Correct
0.95M
0.09M7.21×
bix-38-q1Correct
0.49M
Correct
0.53M
Correct
0.50M
0.48M1.04×
bix-41-q5Correct
0.85M
Correct
0.85M
Correct
0.61M
0.09M8.73×
bix-43-q2Correct
0.92M
Correct
0.86M
Correct
0.89M
1.44M0.62×
bix-43-q4Correct
2.03M
Correct
1.99M
Correct
0.89M
0.93M1.76×
bix-45-q1Incorrect
0.76M
Incorrect
0.69M
Incorrect
0.49M
1.11M0.58×
bix-46-q4Correct
0.42M
Correct
0.68M
Correct
0.64M
0.31M1.89×
bix-47-q3Correct
2.00M
Correct
1.53M
Correct
1.27M
0.39M4.11×
bix-49-q4Correct
1.32M
Correct
3.36M
Correct
1.98M
1.01M2.20×
bix-51-q2Correct
0.66M
Correct
0.75M
Correct
0.71M
0.19M3.75×
bix-51-q8Correct
0.80M
Correct
2.58M
Correct
1.29M
0.25M6.33×
bix-52-q2Correct
1.19M
Correct
2.45M
Correct
1.51M
0.11M15.41×
bix-52-q6Correct
1.47M
Correct
1.37M
Correct
0.90M
0.08M14.91×
bix-52-q7Correct
0.48M
Correct
0.62M
Correct
0.60M
0.06M8.93×
bix-53-q2Correct
1.31M
Correct
1.43M
Correct
0.94M
0.99M1.24×
bix-53-q5Correct
1.60M
Correct
1.72M
Correct
2.00M
0.79M2.26×
bix-54-q7Incorrect
0.58M
Incorrect
0.58M
Incorrect
0.54M
0.12M4.86×
bix-55-q1Correct
0.89M
Correct
0.59M
Correct
0.89M
2.66M0.30×
bix-6-q4Correct
0.63M
Correct
0.92M
Correct
0.80M
0.17M4.69×
bix-61-q2Correct
0.58M
Correct
0.55M
Correct
0.62M
0.39M1.51×
bix-61-q5Incorrect
0.50M
Incorrect
0.53M
Incorrect
0.29M
0.26M1.70×