52.50M tokens · 1.48× Open-ended code
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
60.45M tokens · 1.70× Open-ended code
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
| Configuration | Tokens per complete 50-task run | Relative to Open-ended | Coverage |
|---|---|---|---|
| Open-ended code | 35.51M | 1.00× | 3-replicate mean |
| Previous website Galaxy | 122.41M | 3.45× | 3-replicate mean |
| Final optimized Galaxy | 54.35M | 1.53× | 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
- Shorter successful replies
Return job status, dataset identifiers, relevant results and check summaries. Bound inline text previews; keep full requests, replies and job details for later inspection. Error details remain available.
- Avoid repeated content
Deduplicate output datasets and abbreviate unchanged results and tool schemas within a session. A tool schema is still checked against the live Galaxy interface and can be expanded again.
- Clearer submission formats
Remove repeated parameter descriptions and provide generic templates with actual parameter paths. These describe how to call the selected tool, not how to solve a benchmark question.
- Fewer format-related retries
Correct false parameter-mismatch alarms caused by equivalent repeat-field names or different API input formats. Check custom-tool input/output definitions before creation; keep scientific parameter checks.
- Longer waits between model requests
Use the approved waiting instructions: wait at least five minutes, preferably thirty, rather than repeatedly asking whether an asynchronous call has finished. The Galaxy tool interface already waits for jobs; this change reduces the model's checks of that still-running call.
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.
Lab HF batchMachine-readable results
| Item | R1 | R2 | R3 | Open-ended mean | New Galaxy / Open-ended |
|---|---|---|---|---|---|
| bix-11-q1 | Correct 0.58M | Correct 0.77M | Correct 0.63M | 0.28M | 2.36× |
| bix-11-q2 | Correct 0.72M | Correct 0.51M | Correct 0.57M | 0.40M | 1.52× |
| bix-12-q2 | Correct 1.13M | Correct 2.18M | Correct 1.69M | 0.44M | 3.76× |
| bix-12-q4 | Correct 2.96M | Correct 4.54M | Correct 2.20M | 0.61M | 5.32× |
| bix-12-q5 | Correct 2.85M | Correct 2.39M | Correct 2.55M | 0.50M | 5.15× |
| bix-12-q6 | Correct 1.60M | Correct 0.90M | Correct 2.07M | 0.45M | 3.37× |
| bix-14-q1 | Correct 2.67M | Correct 2.61M | Incorrect 1.92M | 0.50M | 4.81× |
| bix-16-q1 | Correct 2.06M | Correct 0.40M | Correct 0.48M | 0.39M | 2.52× |
| bix-16-q3 | Correct 0.46M | Correct 0.41M | Correct 0.45M | 0.30M | 1.47× |
| bix-16-q4 | Correct 0.62M | Correct 0.35M | Correct 0.56M | 0.28M | 1.79× |
| bix-17-q2 | Correct 1.33M | Correct 1.15M | Correct 1.78M | 0.36M | 4.00× |
| bix-18-q1 | Correct 0.36M | Correct 0.56M | Correct 0.61M | 0.07M | 7.65× |
| bix-18-q3 | Correct 0.49M | Correct 0.32M | Correct 0.41M | 0.08M | 4.97× |
| bix-20-q3 | Correct 1.19M | Correct 1.87M | Correct 1.67M | 0.39M | 4.06× |
| bix-22-q1 | Correct 0.65M | Correct 0.60M | Correct 0.65M | 0.19M | 3.29× |
| bix-22-q4 | Correct 0.90M | Correct 1.79M | Correct 0.83M | 0.18M | 6.45× |
| bix-24-q2 | Correct 1.46M | Correct 3.31M | Correct 1.30M | 1.52M | 1.33× |
| bix-26-q3 | Correct 0.72M | Correct 0.78M | Correct 1.48M | 0.26M | 3.87× |
| bix-26-q5 | Correct 0.75M | Correct 0.82M | Correct 0.75M | 0.43M | 1.80× |
| bix-27-q5 | Correct 0.71M | Correct 0.79M | Correct 0.92M | 0.23M | 3.56× |
| bix-28-q3 | Correct 0.64M | Incorrect 0.62M | Correct 0.72M | 0.52M | 1.27× |
| bix-30-q3 | Correct 2.22M | Correct 1.31M | Correct 0.81M | 0.56M | 2.59× |
| bix-31-q2 | Correct 0.80M | Correct 0.59M | Correct 0.70M | 12.70M | 0.05× |
| bix-32-q2 | Incorrect 1.31M | Correct 1.38M | Incorrect 1.42M | 0.48M | 2.83× |
| bix-34-q2 | Correct 0.93M | Correct 1.85M | Correct 0.82M | 0.21M | 5.64× |
| bix-34-q5 | Correct 0.70M | Correct 0.69M | Correct 0.87M | 0.24M | 3.10× |
| bix-35-q1 | Correct 0.61M | Correct 0.69M | Correct 0.80M | 0.32M | 2.19× |
| bix-35-q2 | Correct 0.58M | Correct 0.63M | Correct 0.57M | 0.59M | 1.00× |
| bix-37-q1 | Correct 0.62M | Correct 0.52M | Correct 0.56M | 0.13M | 4.35× |
| bix-37-q4 | Correct 0.44M | Correct 0.56M | Correct 0.95M | 0.09M | 7.21× |
| bix-38-q1 | Correct 0.49M | Correct 0.53M | Correct 0.50M | 0.48M | 1.04× |
| bix-41-q5 | Correct 0.85M | Correct 0.85M | Correct 0.61M | 0.09M | 8.73× |
| bix-43-q2 | Correct 0.92M | Correct 0.86M | Correct 0.89M | 1.44M | 0.62× |
| bix-43-q4 | Correct 2.03M | Correct 1.99M | Correct 0.89M | 0.93M | 1.76× |
| bix-45-q1 | Incorrect 0.76M | Incorrect 0.69M | Incorrect 0.49M | 1.11M | 0.58× |
| bix-46-q4 | Correct 0.42M | Correct 0.68M | Correct 0.64M | 0.31M | 1.89× |
| bix-47-q3 | Correct 2.00M | Correct 1.53M | Correct 1.27M | 0.39M | 4.11× |
| bix-49-q4 | Correct 1.32M | Correct 3.36M | Correct 1.98M | 1.01M | 2.20× |
| bix-51-q2 | Correct 0.66M | Correct 0.75M | Correct 0.71M | 0.19M | 3.75× |
| bix-51-q8 | Correct 0.80M | Correct 2.58M | Correct 1.29M | 0.25M | 6.33× |
| bix-52-q2 | Correct 1.19M | Correct 2.45M | Correct 1.51M | 0.11M | 15.41× |
| bix-52-q6 | Correct 1.47M | Correct 1.37M | Correct 0.90M | 0.08M | 14.91× |
| bix-52-q7 | Correct 0.48M | Correct 0.62M | Correct 0.60M | 0.06M | 8.93× |
| bix-53-q2 | Correct 1.31M | Correct 1.43M | Correct 0.94M | 0.99M | 1.24× |
| bix-53-q5 | Correct 1.60M | Correct 1.72M | Correct 2.00M | 0.79M | 2.26× |
| bix-54-q7 | Incorrect 0.58M | Incorrect 0.58M | Incorrect 0.54M | 0.12M | 4.86× |
| bix-55-q1 | Correct 0.89M | Correct 0.59M | Correct 0.89M | 2.66M | 0.30× |
| bix-6-q4 | Correct 0.63M | Correct 0.92M | Correct 0.80M | 0.17M | 4.69× |
| bix-61-q2 | Correct 0.58M | Correct 0.55M | Correct 0.62M | 0.39M | 1.51× |
| bix-61-q5 | Incorrect 0.50M | Incorrect 0.53M | Incorrect 0.29M | 0.26M | 1.70× |