GPT-5.4 Pro ekes slim gains on MineBench

The cost-to-performance ratio is the real headline here: GPT-5.4-Pro's spatial reasoning gains are real but hard to justify at $29 a pop when standard 5.4 runs a full benchmark sweep for the same price.

• –MineBench tasks models to return raw 3D block coordinates as JSON from text prompts — no images, no visual feedback — making it a genuine test of internal spatial modeling, not prompt mimicry
• –Builds averaged 56 minutes each (max 76 min), reflecting genuinely complex generation; these aren't toy evals
• –Creator flags a key benchmark design flaw: the system prompt may not push Pro-tier models to use their extended reasoning budgets, potentially understating the gap
• –At $435 for 15 API calls, community benchmarks like MineBench expose how fast frontier model costs spiral — and why reproducible third-party evals are increasingly reliant on crowdfunding
• –MineBench uses a Glicko-style rating with arena community voting, giving it more statistical rigor than most hobby benchmarks; TechCrunch covered it as a notable independent eval effort