Reported model tokens and estimated API cost
The totals come directly from the final usage record in each available Codex event log.
240/240 usage records1.71B recorded tokens$967.28 estimated cost
Experiment totals
- Coverage: 240 of 240 selected runs have a final token-usage record. Totals and averages use those directly recorded values; no missing values are estimated.
- Token accounting: total tokens are input plus output. Cached input is part of input; reasoning tokens are part of output and are not counted or charged twice.
- wf007 update: the 24 ERR12833637 runs replace the earlier wf007 dataset. The other 216 selected runs are unchanged.
- OpenAI cost: API-equivalent estimates use Fast rates per 1M uncached input / cached input / output tokens: GPT-5.5 $12.50 / $1.25 / $75, Sol $8 / $0.80 / $40, and Luna $0.40 / $0.04 / $2.40.
- DeepSeek cost: estimates use the weekday UTC schedule in effect for these runs: peak $1.32 / $0.044 / $3.96 and off-peak $0.66 / $0.022 / $1.98 per 1M uncached input / cached input / output tokens. For a run crossing a rate boundary, tokens are distributed evenly over its duration.
Prices checked October 11, 2026: OpenAI Fast mode · DeepSeek pricing. Source traces: Hugging Face IWC collection.
| Agent and execution | Total | Input | Cached input | Uncached input | Output | Reasoning output | Estimated costAPI-equivalent | Cost / recorded run | Tokens / recorded run |
|---|---|---|---|---|---|---|---|---|---|
| GPT-5.5Open-ended code30/30 run usage records | 215.6M | 214.9M | 211.1M | 3.8M | 720.9K | 193.6K | $365.54 | $12.18 | 7.2M |
| GPT-5.5Galaxy-API code30/30 run usage records | 83.1M | 82.6M | 78.2M | 4.4M | 506.5K | 143.8K | $191.00 | $6.37 | 2.8M |
| GPT-5.6 SolOpen-ended code30/30 run usage records | 111.4M | 110.8M | 107.4M | 3.3M | 579.0K | 229.3K | $135.87 | $4.53 | 3.7M |
| GPT-5.6 SolGalaxy-API code30/30 run usage records | 196.3M | 195.6M | 190.5M | 5.2M | 606.1K | 223.6K | $218.02 | $7.27 | 6.5M |
| GPT-5.6 LunaOpen-ended code30/30 run usage records | 310.3M | 309.0M | 301.8M | 7.2M | 1.3M | 638.4K | $18.03 | $0.60 | 10.3M |
| GPT-5.6 LunaGalaxy-API code30/30 run usage records | 330.6M | 329.4M | 318.8M | 10.5M | 1.2M | 700.4K | $19.95 | $0.67 | 11.0M |
| Codex + DeepSeek V4 ProOpen-ended code30/30 run usage records | 234.9M | 233.9M | 231.9M | 2.1M | 1.0M | 346.5K | $8.48 | $0.28 | 7.8M |
| Codex + DeepSeek V4 ProGalaxy-API code30/30 run usage records | 226.0M | 224.8M | 220.4M | 4.5M | 1.2M | 435.3K | $10.38 | $0.35 | 7.5M |
wf007 · ERR12833637: tokens and cost
Three-run totals for the same input library in each condition. The ratio is Galaxy-API code tokens divided by Open-ended code tokens. Costs are API-equivalent estimates.
| Model | Open-ended code tokens | Galaxy-API code tokens | Galaxy / Open-ended | Open-ended code cost | Galaxy-API code cost |
|---|---|---|---|---|---|
| GPT-5.5 | 8.9M | 6.8M | 0.761× | $20.57 | $18.92 |
| GPT-5.6 Sol | 25.1M | 37.9M | 1.509× | $28.49 | $40.33 |
| GPT-5.6 Luna | 60.0M | 32.7M | 0.545× | $3.42 | $1.96 |
| Codex + DeepSeek V4 Pro | 51.0M | 6.7M | 0.132× | $1.67 | $0.37 |