- Input
- 443.4M base
- Cached input
- 423.4M base
- Uncached input
- 20.0M base
- Output
- 3.3M base
- Reasoning output
- 813.1K base
Condition totals
Total tokens are computed as input tokens plus output tokens from each replicate's usage record. Percent changes are relative to Galaxy-API code with skills.
- Input
- 499.6M +12.7%
- Cached input
- 481.2M +13.7%
- Uncached input
- 18.4M -8.4%
- Output
- 3.3M -2.2%
- Reasoning output
- 855.0K +5.1%
- Input
- 80.4M -81.9%
- Cached input
- 72.4M -82.9%
- Uncached input
- 8.1M -59.8%
- Output
- 1.1M -66.8%
- Reasoning output
- 329.0K -59.5%
- Input
- 67.3M -84.8%
- Cached input
- 60.0M -85.8%
- Uncached input
- 7.3M -63.7%
- Output
- 941.0K -71.9%
- Reasoning output
- 272.9K -66.4%
Why token use differs
The differences mostly come from platform context, not from answer length.
Galaxy conditions are much larger than anycode
Galaxy-API code includes Galaxy API interaction, history inspection, job/tool evidence, UDT or tool-selection context, and longer trace reasoning. Open-ended code does local file/code work and avoids that Galaxy evidence chain, so both anycode conditions use more than 80% fewer total tokens than Galaxy-API code with skills.
Galaxy without skills is higher than the base
Galaxy-API code without skills uses 12.3% more total tokens than Galaxy-API code with skills. Almost all of that increase is input context: cached input is 13.3% higher, while uncached input is 7.6% lower and output is essentially flat. The with-skills condition includes `galaxy-token-efficient-operations`, which tells the agent to save bulky Galaxy evidence to trace files and keep transcript output compact. Without that token-saving Galaxy skill, the agent repeats more Galaxy/tool/history context in the conversation.
Open-ended code with skills is still far below the base
Open-ended code with skills uses 81.4% fewer total tokens than the Galaxy base because it does not carry Galaxy histories, API schemas, or tool/job evidence. It is higher than anycode without skills because skills add setup guidance and domain checks, which increase input and reasoning tokens even without Galaxy.
Open-ended code without skills is the smallest
Open-ended code without skills has the lowest token usage: 84.4% below the Galaxy base. It has the shortest context stack: no Galaxy context and no skill instructions. That saves tokens, but the accuracy page shows it also has the lowest accuracy among the four conditions.
Replicate-level usage
Each replicate column is compared to the same replicate in Galaxy-API code with skills.
| Condition | Total | replicate 1 | replicate 2 | replicate 3 | Cached input | Uncached input | Output | Reasoning output |
|---|---|---|---|---|---|---|---|---|
| Galaxy-API code with skills | 446.7M base | 144.7M base 50 items | 142.8M base 50 items | 159.2M base 50 items | 423.4M base | 20.0M base | 3.3M base | 813.1K base |
| Galaxy-API code without skills | 502.8M +12.6% | 177.0M +22.3% 50 items | 179.3M +25.5% 50 items | 146.6M -7.9% 50 items | 481.2M +13.7% | 18.4M -8.4% | 3.3M -2.2% | 855.0K +5.1% |
| Open-ended code with skills | 81.5M -81.8% | 24.7M -83.0% 50 items | 28.2M -80.2% 50 items | 28.6M -82.0% 50 items | 72.4M -82.9% | 8.1M -59.8% | 1.1M -66.8% | 329.0K -59.5% |
| Open-ended code without skills | 68.2M -84.7% | 23.4M -83.8% 50 items | 21.1M -85.2% 50 items | 23.7M -85.1% 50 items | 60.0M -85.8% | 7.3M -63.7% | 941.0K -71.9% | 272.9K -66.4% |