Opus, Neon trail GPT on RedstoneBench
Benchmark results for Anthropic's Claude Opus 5.5 and Google's Gemini 4 Neon were added to RedstoneBench, an automated Minecraft circuit evaluation suite. Both frontier models struggled with spatial logic and mechanical timing, failing to match OpenAI's GPT-6 series in circuit compactness and high-complexity piston tasks.
Real-world physical and game-engine environments expose the persistent blind spots of modern reasoning models in 3D spatial planning. While both frontier models can write basic logic gates, complex sequential mechanisms like piston doors reveal a severe gap between benchmark coding prowess and true spatial reasoning.
- –Claude Opus 5.5 completely failed both the 3x3 and 4x4 piston door challenges, stalling on multi-tick block retractors and sequential timings.
- –Gemini 4 Neon completed all 16 tasks via its native agy runner, but its solutions were heavily unoptimized—its 3x3 door consumed 2,184 blocks compared to GPT-6.2 Astra's 88.
- –GPT-6.2 Astra remains the clear leader in spatial-logical compression, outperforming both competitors by an order of magnitude on complex mechanical and memory challenges.
- –RedstoneBench highlights the value of deterministic, tick-by-tick programmatic verification over subjective LLM judges for measuring agentic problem-solving.
DISCOVERED
3h ago
2026-11-20
PUBLISHED
3h ago
2026-11-20
RELEVANCE
AUTHOR
AdamHoltererer