Been using GPT-5.4 these past few days, and the reason is simple—check out the chart below USAMO 2026 Accuracy Test (Cost per million tokens): GPT-5.4 (xhigh) ranked #1 with 95.24% accuracy, costing just $5.15 Claude Opus 4.6 only hit 47.02%, performing way below expectations DeepSeek-R1-0528 came in with 30.06% accuracy + an ultra-low cost of just $0.23, making it the biggest dark horse! Conclusion: For ultimate performance → GPT-5.4 For ultimate cost-effectiveness → DeepSeek-R1 Coding folks, which model are you mainly using now?