claude-opus-4-8
claude-opus-4-8 (Anthropic) scored 79.1% composite across 87 tasks. This run covers Australian accounting, Animation, Auth screens, Dashboards, Australian law, Marketing pages, Programming, SVG and graphics, svg-scene, UI components, Websites. Published 2026-07-21.
Agent harnessMedium reasoning
Composite
79.1%
95% CI 71.2% - 86.4%
Est. cost
$3.45
USD
Tokens
181,820
generation
Wall clock
8 min
end to end
Gate pass
-
validated
Separation
-
good vs bad
By domain
Composite score per domain, weakest first. Judge is the vision model’s read, shown for the visual domains.
Programming
6 tasks
33.3%
composite
Marketing page
1 task
37.5%
judge
37.5%
composite
Websites
15 tasks
53.9%
composite
svg-scene
1 task
64.6%
judge
64.6%
composite
Animation
2 tasks
70.8%
judge
70.8%
composite
Auth screen
1 task
77.1%
judge
77.1%
composite
Dashboard
1 task
85.4%
judge
85.4%
composite
SVG and graphics
13 tasks
86.6%
composite
Australian accounting
23 tasks
89.2%
composite
UI components
19 tasks
94.4%
composite
Australian law
5 tasks
100.0%
composite
Outputs
The actual rendered output. Open any tile to view it in a popup, or compare the same task across every model.
53 results
![claude-opus-4-8 SVG and graphics output - Bar chart with exact proportional heights for [4, 10, 6, 8, 12]](/lab-thumbs/opus48-medium/DIAGR-0001.jpg)
![claude-opus-4-8 SVG and graphics output - Tiny bar chart for values [3, 7, 5, 9]](/lab-thumbs/opus48-medium/SVG-0005.jpg)
Objective tasks
Programming, Australian legal and accounting, graded by execution. 25 of 34 scored a perfect 100.0%; the rest are below. Open the answer in a popup, or compare it across every model.