GPT-6 Astra
Key numbers
- Panel score
- 63rank 4 of 10
- Cost
- $4.46API list price
- Wall time
- 14 min
- Model time
- 11 min81% of wall time
- tok/s
- 36.2
- Tokens
- 2.6MOutput: 25k
- Cache hit
- 97%
- API calls
- 35
- Tool calls
- 34
- Renderer
- Intel UHD 630 (GPU)~99 ms / frame
- H1–H6
- Pass
- Cost per point
- $0.071
How the run ended
Finished cleanly with a passing video and a report. The audio track is silent.
Blind panel
Scores of the three judges in every panel this video was part of.
| Panel | Letter | Anthropic | OpenAI | Median | Spread | Rank | |
|---|---|---|---|---|---|---|---|
| 14 Nov 2026 11 videos | C | 58 | 63 | 68 | 63 | 10 | 4 / 11 |
| 13 Nov 2026 10 videos | F | 64 | 61 | 65 | 64 | 4 | 4 / 10 |
| 11 Nov 2026 9 videos | B | 55 | 58 | 77 | 58 | 22 | 4 / 9 |
Sub-scores and the judges’ one-line notes from the panel of 14 Nov 2026.
| judge | Landmarks / 25 | City look / 25 | Camera & drone / 30 | Life / 10 | Overall / 10 | Total |
|---|---|---|---|---|---|---|
| Anthropic | 14 | 13 | 19 | 7 | 5 | 58 |
| OpenAI | 13 | 14 | 24 | 6 | 6 | 63 |
| 18 | 16 | 22 | 6 | 6 | 68 |
Judges’ notes
- AnthropicDrone readable throughout but flat shading, odd black-square people, and two frames blocked by black slabs.
- OpenAIA clear drone chase and populated bridge compensate partly for generic architecture and sparse landscaping.
- GoogleClean procedural colors and a very readable quadcopter, but simpler blocky buildings and minor edge clipping at the bridge tower and finale.
Time and speed
Wall time includes rendering and encoding. Model time is the sum of all model calls.
- Model generating
- Tools & waiting
| Wall time | 14 min |
|---|---|
| Model time | 11 min |
| Tools & waiting | 2.7 min |
| Model share of wall time | 81% |
| tok/s | 36.2 |
| Model calls | 35 |
Tokens and cost
Tokens summed over every model response. Tokens multiplied by the provider’s price.
| Uncached input | 69,270 |
|---|---|
| Cache read | 2,541,952 |
| Cache write | 0 |
| Output | 24,527of which 3,129 reasoning |
| Total | 2,635,749 |
| Cache-hit ratio | 97.4% |
| Cost | $4.46API list price |
| Output tokens per point | 389 |
| Points per $ | 14.1 |
Tools used
Tool calls by name, as the harness logged them. Names differ between harnesses.
Rendering
| Renderer | Intel UHD 630 (GPU) |
|---|---|
| Final render | 3.0 min |
| Per frame | 99 ms (approximate) |
Hard requirements H1–H6
Checked independently from the delivered MP4 with ffprobe.
| H1 | h264 High yuv420p | Pass | |
|---|---|---|---|
| H1 | faststart | Pass | |
| H2 | 1080×1920 SAR1:1 | Pass | 1080x1920 SAR 1:1 |
| H3 | 30 fps | Pass | 30 fps |
| H4 | 55-60 s | Pass | 58.000 s |
| H5 | peak≤ 20Mbps, <200MB | Pass | peak 14.80 Mbps over any 1 s, 69.8 MB |
| H6 | AAC-LC 48k stereo | Pass |
From the agent’s report
Short, verbatim quotes from the REPORT.md the agent wrote itself.
This is a simplified geometric interpretation, not an architectural reconstruction.
Contact sheet and route map
