prague-bench

← All runs

Claude Opus 5.5

FinishedAnthropicClaude CodeEffort: high11 Nov 2026spec 666c07f57d6e

Video preview · 540 × 960 · 5.7 MB

Key numbers

Panel score
81rank 2 of 10
Cost
$15.41API list price
Wall time
1 h 7 min
Model time
56 min83% of wall time
tok/s
105.3
Tokens
23.2MOutput: 352k
Cache hit
97%
API calls
76
Tool calls
77
Renderer
Intel UHD 630 (GPU)
H1–H6
Pass
Cost per point
$0.190

How the run ended

Finished cleanly with a passing video, a procedural soundtrack and a full report.

Blind panel

Scores of the three judges in every panel this video was part of.

Blind panel
PanelLetterAnthropicOpenAIGoogleMedianSpreadRank
14 Nov 2026 11 videosK85817581102 / 11
13 Nov 2026 10 videosA7983878382 / 10
11 Nov 2026 9 videosA7983808042 / 9

Sub-scores and the judges’ one-line notes from the panel of 14 Nov 2026.

Sub-scores and the judges’ one-line notes from the panel of 14 Nov 2026.
judgeLandmarks / 25City look / 25Camera & drone / 30Life / 10Overall / 10Total
Anthropic2122249985
OpenAI2021248881
Google2121197775

Judges’ notes

  • AnthropicNear-top quality with rich lighting, dense crowds, recognisable cathedral, dome and towers; one gate frame half-blocked.
  • OpenAIRich textures, substantial crowds and distinctive monuments impress, but tight gate framing and heavy banking reduce clarity.
  • GoogleRich warm Prague architecture and readable drone throughout most of the flight, but penalized for a dark tower-gate occlusion at 36.2 s.

Time and speed

Wall time includes rendering and encoding. Model time is the sum of all model calls.

  • Model generating
  • Tools & waiting
  1. Wall time1 h 7 min83% model
Time and speed
Wall time1 h 7 min
Model time56 min
Tools & waiting12 min
Model share of wall time83%
tok/s105.3
Model calls76

Tokens and cost

Tokens summed over every model response. Tokens multiplied by the provider’s price.

Tokens and cost
Uncached input158
Cache read22,010,262
Cache write793,797
Output351,898
Total23,156,115
Cache-hit ratio96.5%
Cost$15.41API list price
Output tokens per point4,344
Points per $5.3

Tools used

Tool calls by name, as the harness logged them. Names differ between harnesses.

  1. Bash47
  2. Write20
  3. Read10

Rendering

Rendering
RendererIntel UHD 630 (GPU)
Final render4.9 min
Per frame–

Hard requirements H1–H6

Checked independently from the delivered MP4 with ffprobe.

Hard requirements H1–H6
H1h264 High yuv420pPass
H1faststartPass
H21080×1920 SAR1:1Pass1080x1920 SAR 1:1
H330 fpsPass30 fps
H455-60 sPass58.000 s
H5peak≤ 20Mbps, <200MBPasspeak 14.98 Mbps over any 1 s, 83.2 MB
H6AAC-LC 48k stereoPass

From the agent’s report

Short, verbatim quotes from the REPORT.md the agent wrote itself.

The hairpin pivot is quick, about 2 s at roughly 12 m/s. The camera rises and re-frames, but the move is still energetic.

A CPU-only (SwiftShader) final render would take hours, at about 15 s per frame.

Contact sheet and route map