I found benchmark results on Reddit and brought them here.
I've briefly summarized the original content.
'MineBench' is an open-source benchmark that tests architectural quality by giving AI a prompt and having it extract Minecraft-style 3D structure coordinates (JSON). This is data comparing the newly released Claude Fable 5 with the existing Opus 4.8.
1. Speed and Cost (Based on 15 Build Tests)
2. Build Quality (Subjective Evaluation)
The detail is incredible. The size doesn't feel as massively expanded as the official benchmark scores suggest, but the obsession with detail is no joke. For example, when asked to create an 'Arcade Machine,' it's reportedly the first model to accurately implement not just the Pac-Man layout on the screen, but also the score board and the '1UP' display.
However, it interprets prompts somewhat conservatively. Since the scale itself hasn't grown grandiosely, some builds apparently look more impressive with Opus 4.8 results at first glance. Overall, it feels like it creates only what's needed efficiently and compactly.
Inferring Coding Performance According to test users, they speculate that Fable's reputation as a beast in coding may be due to this characteristic of producing code that is intuitive and clean without any unnecessary frills.
3. Tips for Anyone Testing (VoxelBench Team Feedback)
Since Fable 5 tends to create structures somewhat small by default, adding the two lines below to the prompt template reportedly results in a dramatic increase in quality and size.
To summarize, Fable 5 can be described as "a tightly packed style that costs a bit more but is fast, doesn't boast, and obsesses extremely over details".
After seeing the benchmark results, Fable5 and Opus4.8 feel quite different, you know?
As for the cost part, they say it's not as high as expected, but I think we should withhold judgment on that for now.
