GPT-5.4 Solves Novel Math via Obscure Preprint

🔑 Enhanced Key Takeaways

• •GPT-5.4 achieved 38% on FrontierMath Tier 4 across 10 runs, including solving a 20-year unsolved problem praised by mathematicians as 'very nice, clean, and feels almost human'.[1][3]
• •On OSWorld desktop navigation benchmark, GPT-5.4 scored 75%, surpassing the human average of 72.4%.[5]
• •GPT-5.4 reduces hallucinations with individual claims 33% less likely to be false and full responses 18% less likely to contain errors compared to GPT-5.2.[6]
• •Introduces native computer-use capabilities as OpenAI's first general-purpose model for autonomous desktop, browser, and software navigation.[4][5][6]

🛠️ Technical Deep Dive

• •Native computer-use capabilities include new image input levels: 'original' up to 10.24M pixels (6000px max dimension) and updated 'high' up to 2.56M pixels (2048px max).[4]
• •Reasoning effort levels such as 'xhigh' used for most benchmarks; at 'none', still outperforms GPT-5.2 on latency-sensitive tasks like τ²-bench Telecom (64.3% vs 57.2%).[4]
• •Fast mode delivers up to 1.5x faster token velocity; API priority processing equivalent.[4]
• •New tool search feature cuts token usage by 47%; advanced steerability generates upfront thinking plans adjustable mid-response.[2][5]

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline