CS2 reasoning tokens
72 full traces. One question.
Give me the routes + actions of each player.
- Full traces
- 72
- Models
- 3
- Situations
- 5
- Repetitions
- 5 per pair
Three models received five fixed situations from a professional Dust II recording, five times each. Every returned response is here, including the full thinking text, exact prompt, player routes/actions and timing.
75 calls produced 72 responses. One timeout and two provider rate limits produced no trace. Two answers vary from the requested format and are included unchanged. No CS2 simulation was run.
Request timing
End-to-end latency for returned responses, including provider waiting and generation.
| Model | Responses | Median time | Range |
|---|---|---|---|
| DeepSeek V4 Pro 0813 | 22/25 | 485.88 s | 100.67–923.22 s |
| Qwen3.8 2.4T A95B | 25/25 | 64.04 s | 44.40–89.19 s |
| Kimi K3.1 Pro | 25/25 | 23.64 s | 11.31–103.69 s |
All traces
72 traces
| Situation | Model | Run | Time | Thinking tokens | Output format | Trace |
|---|---|---|---|---|---|---|
| Long approach | DeepSeek V4 Pro 0813 | 3/5 | 452.19 s | 6674 | Recorded | Read full trace ↗ |
| Long approach | DeepSeek V4 Pro 0813 | 4/5 | 375.30 s | 5748 | Recorded | Read full trace ↗ |
| Long approach | DeepSeek V4 Pro 0813 | 5/5 | 467.30 s | 7310 | Recorded | Read full trace ↗ |
| Long approach | Qwen3.8 2.4T A95B | 1/5 | 80.71 s | 2386 | Recorded | Read full trace ↗ |
| Long approach | Qwen3.8 2.4T A95B | 2/5 | 70.34 s | 1915 | Recorded | Read full trace ↗ |
| Long approach | Qwen3.8 2.4T A95B | 3/5 | 54.17 s | 1442 | Recorded | Read full trace ↗ |
| Long approach | Qwen3.8 2.4T A95B | 4/5 | 49.00 s | 1319 | Recorded | Read full trace ↗ |
| Long approach | Qwen3.8 2.4T A95B | 5/5 | 61.48 s | 1955 | Recorded | Read full trace ↗ |
| Long approach | Kimi K3.1 Pro | 1/5 | 23.31 s | 2263 | Recorded | Read full trace ↗ |
| Long approach | Kimi K3.1 Pro | 2/5 | 22.03 s | 2176 | Recorded | Read full trace ↗ |
| Long approach | Kimi K3.1 Pro | 3/5 | 27.14 s | 2644 | Recorded | Read full trace ↗ |
| Long approach | Kimi K3.1 Pro | 4/5 | 35.40 s | 3516 | Format variation | Read full trace ↗ |
| Long approach | Kimi K3.1 Pro | 5/5 | 43.21 s | 2135 | Recorded | Read full trace ↗ |
| Mid-round disadvantage | DeepSeek V4 Pro 0813 | 1/5 | 683.53 s | 10617 | Recorded | Read full trace ↗ |
| Mid-round disadvantage | DeepSeek V4 Pro 0813 | 2/5 | 749.70 s | 12128 | Recorded | Read full trace ↗ |
| Mid-round disadvantage | DeepSeek V4 Pro 0813 | 3/5 | 599.16 s | 9890 | Recorded | Read full trace ↗ |
| Mid-round disadvantage | DeepSeek V4 Pro 0813 | 4/5 | 923.22 s | 13576 | Recorded | Read full trace ↗ |
| Mid-round disadvantage | DeepSeek V4 Pro 0813 | 5/5 | 703.68 s | 10496 | Recorded | Read full trace ↗ |
| Mid-round disadvantage | Qwen3.8 2.4T A95B | 1/5 | 61.60 s | 1377 | Recorded | Read full trace ↗ |
| Mid-round disadvantage | Qwen3.8 2.4T A95B | 2/5 | 82.24 s | 2255 | Recorded | Read full trace ↗ |
| Mid-round disadvantage | Qwen3.8 2.4T A95B | 3/5 | 89.19 s | 2632 | Recorded | Read full trace ↗ |
| Mid-round disadvantage | Qwen3.8 2.4T A95B | 4/5 | 75.88 s | 2390 | Recorded | Read full trace ↗ |
| Mid-round disadvantage | Qwen3.8 2.4T A95B | 5/5 | 67.47 s | 2095 | Recorded | Read full trace ↗ |
| Mid-round disadvantage | Kimi K3.1 Pro | 1/5 | 26.06 s | 3068 | Recorded | Read full trace ↗ |
| Mid-round disadvantage | Kimi K3.1 Pro | 2/5 | 18.49 s | 1784 | Recorded | Read full trace ↗ |
| Mid-round disadvantage | Kimi K3.1 Pro | 3/5 | 23.64 s | 1739 | Recorded | Read full trace ↗ |
| Mid-round disadvantage | Kimi K3.1 Pro | 4/5 | 39.86 s | 1687 | Recorded | Read full trace ↗ |
| Mid-round disadvantage | Kimi K3.1 Pro | 5/5 | 22.23 s | 2347 | Recorded | Read full trace ↗ |
| B post-plant | DeepSeek V4 Pro 0813 | 1/5 | 249.35 s | 3495 | Recorded | Read full trace ↗ |
| B post-plant | DeepSeek V4 Pro 0813 | 2/5 | 100.67 s | 1348 | Recorded | Read full trace ↗ |
| B post-plant | DeepSeek V4 Pro 0813 | 3/5 | 153.79 s | 2034 | Recorded | Read full trace ↗ |
| B post-plant | DeepSeek V4 Pro 0813 | 4/5 | 205.42 s | 2981 | Recorded | Read full trace ↗ |
| B post-plant | DeepSeek V4 Pro 0813 | 5/5 | 162.98 s | 2524 | Recorded | Read full trace ↗ |
| B post-plant | Qwen3.8 2.4T A95B | 1/5 | 67.24 s | 1753 | Recorded | Read full trace ↗ |
| B post-plant | Qwen3.8 2.4T A95B | 2/5 | 64.47 s | 1430 | Recorded | Read full trace ↗ |
| B post-plant | Qwen3.8 2.4T A95B | 3/5 | 72.94 s | 1501 | Recorded | Read full trace ↗ |
| B post-plant | Qwen3.8 2.4T A95B | 4/5 | 58.36 s | 1504 | Recorded | Read full trace ↗ |
| B post-plant | Qwen3.8 2.4T A95B | 5/5 | 61.67 s | 1565 | Recorded | Read full trace ↗ |
| B post-plant | Kimi K3.1 Pro | 1/5 | 36.72 s | 546 | Recorded | Read full trace ↗ |
| B post-plant | Kimi K3.1 Pro | 2/5 | 14.65 s | 1470 | Recorded | Read full trace ↗ |
| B post-plant | Kimi K3.1 Pro | 3/5 | 15.74 s | 983 | Recorded | Read full trace ↗ |
| B post-plant | Kimi K3.1 Pro | 4/5 | 15.98 s | 1562 | Recorded | Read full trace ↗ |
| B post-plant | Kimi K3.1 Pro | 5/5 | 11.31 s | 971 | Recorded | Read full trace ↗ |
| Protect the advantage | DeepSeek V4 Pro 0813 | 1/5 | 515.87 s | 7678 | Recorded | Read full trace ↗ |
| Protect the advantage | DeepSeek V4 Pro 0813 | 2/5 | 523.44 s | 7492 | Format variation | Read full trace ↗ |
| Protect the advantage | DeepSeek V4 Pro 0813 | 3/5 | 690.21 s | 10564 | Recorded | Read full trace ↗ |
| Protect the advantage | DeepSeek V4 Pro 0813 | 4/5 | 433.61 s | 6309 | Recorded | Read full trace ↗ |
| Protect the advantage | DeepSeek V4 Pro 0813 | 5/5 | 506.90 s | 7524 | Recorded | Read full trace ↗ |
| Protect the advantage | Qwen3.8 2.4T A95B | 1/5 | 64.04 s | 1701 | Recorded | Read full trace ↗ |
| Protect the advantage | Qwen3.8 2.4T A95B | 2/5 | 81.20 s | 2195 | Recorded | Read full trace ↗ |
| Protect the advantage | Qwen3.8 2.4T A95B | 3/5 | 56.53 s | 1261 | Recorded | Read full trace ↗ |
| Protect the advantage | Qwen3.8 2.4T A95B | 4/5 | 64.17 s | 1918 | Recorded | Read full trace ↗ |
| Protect the advantage | Qwen3.8 2.4T A95B | 5/5 | 55.72 s | 1554 | Recorded | Read full trace ↗ |
| Protect the advantage | Kimi K3.1 Pro | 1/5 | 33.87 s | 3162 | Recorded | Read full trace ↗ |
| Protect the advantage | Kimi K3.1 Pro | 2/5 | 25.80 s | 3070 | Recorded | Read full trace ↗ |
| Protect the advantage | Kimi K3.1 Pro | 3/5 | 103.69 s | 1890 | Recorded | Read full trace ↗ |
| Protect the advantage | Kimi K3.1 Pro | 4/5 | 37.86 s | 3159 | Recorded | Read full trace ↗ |
| Protect the advantage | Kimi K3.1 Pro | 5/5 | 29.17 s | 2514 | Recorded | Read full trace ↗ |
| Late commitment | DeepSeek V4 Pro 0813 | 1/5 | 504.46 s | 8204 | Recorded | Read full trace ↗ |
| Late commitment | DeepSeek V4 Pro 0813 | 2/5 | 373.34 s | 5615 | Recorded | Read full trace ↗ |
| Late commitment | DeepSeek V4 Pro 0813 | 3/5 | 439.76 s | 6878 | Recorded | Read full trace ↗ |
| Late commitment | DeepSeek V4 Pro 0813 | 4/5 | 613.69 s | 9495 | Recorded | Read full trace ↗ |
| Late commitment | Qwen3.8 2.4T A95B | 1/5 | 66.39 s | 1565 | Recorded | Read full trace ↗ |
| Late commitment | Qwen3.8 2.4T A95B | 2/5 | 58.05 s | 1432 | Recorded | Read full trace ↗ |
| Late commitment | Qwen3.8 2.4T A95B | 3/5 | 62.98 s | 1859 | Recorded | Read full trace ↗ |
| Late commitment | Qwen3.8 2.4T A95B | 4/5 | 61.72 s | 1639 | Recorded | Read full trace ↗ |
| Late commitment | Qwen3.8 2.4T A95B | 5/5 | 44.40 s | 1098 | Recorded | Read full trace ↗ |
| Late commitment | Kimi K3.1 Pro | 1/5 | 28.76 s | 1826 | Recorded | Read full trace ↗ |
| Late commitment | Kimi K3.1 Pro | 2/5 | 21.69 s | 2167 | Recorded | Read full trace ↗ |
| Late commitment | Kimi K3.1 Pro | 3/5 | 22.39 s | 1735 | Recorded | Read full trace ↗ |
| Late commitment | Kimi K3.1 Pro | 4/5 | 19.62 s | 1724 | Recorded | Read full trace ↗ |
| Late commitment | Kimi K3.1 Pro | 5/5 | 20.75 s | 2089 | Recorded | Read full trace ↗ |
Collection notes
Only the Terrorists’ combined observations were supplied. Enemy sightings retain their age. Match names, later outcomes and future movements were excluded from prompts. The returned thinking is provider-exposed text, not a guarantee that all internal reasoning is visible.
DeepSeek and Kimi used high reasoning effort; Qwen used medium because that endpoint does not support high. The requested temperature was 0.7, with a 16,384-token output limit; provider thinking modes may override sampling settings. The first 14 attempts used three workers, then the remaining 61 used nine. There were no retries or follow-up model calls.