GPT-6 Astra on LVBench
81.6%Rank #2 of 9 on LVBench
- Coverage
- Full (1549/1549)
- Cost per task
- $3.783
- Median latency
- 28.5s
- Tokens per task
- 377,893 in · 75 out
Run setting: 1 fps, 1500-frame cap (103 videos thinned).
Tokens and latency
Tokens per call
| Token type | Mean | Median | 90th pct. |
|---|---|---|---|
| Input | 377,893 | 376,569 | 376,639 |
| of which cached | 351,827 | 376,528 | 376,528 |
| Output | 75 | 21 | 176 |
| of which reasoning | 58 | 12 | 156 |
Over 1,612 calls that returned an answer, including questions asked again after a failed attempt.
Latency
- Median
- 28.5s
- Mean
- 52.7s
- 90th percentile
- 105.1s
- 99th percentile
- 351.8s
Per call, end to end including file upload and processing.