OpenAI: o4 Mini High

A peer-reviewed study published January 7, 2026 in Frontiers in Medicine evaluated OpenAI o4-mini-high alongside Claude 4 Opus, Gemini 2.5 Pro, and Qwen 3 on the 200-item New England Journal of Medicine Image Challenge. The variant achieved the highest overall accuracy at 94%, with consistently strong performance acros An error analysis found that 83.3% of o4-mini-high's mistakes reflected lapses in diagnostic logic rather than input processing, and simple prompting techniques including chain-of-thought and few-shot learning corrected over half of these errors. While the evaluation is domain-specific and does not constitute release o

The PricePerToken MedQA leaderboard, dated September 14, 2026, lists o4 Mini High as the top-scoring model on the MedQA benchmark with a 95.2% accuracy, ahead of Gemini 2.5 Pro at 94.6% and DeepSeek R1 at 92.1%. The leaderboard evaluates 23 models on USMLE-style medical question answering using data from LayerLens. Pri As a third-party aggregator leaderboard with a single-benchmark scope, this entry provides a current snapshot of o4 Mini High's performance on a medical reasoning evaluation but no broader behavioral, release, or API-change information. The pricing displayed reflects an upstream aggregator (OpenRouter) rather than firs

The Benchgen model card explicitly identifies o4-mini (high) as the high reasoning-effort configuration of OpenAI's April 2025 compact reasoning model, describing "high" as the maximum-accuracy setting that allocates the largest internal reasoning token budget. It reports variant-specific benchmark scores of 80.2% on L The same model card flags trade-offs for the high-effort variant, including longer latency from expanded reasoning token budgets, making it unsuitable for low-latency high-throughput pipelines, and the absence of open weights. Recommended use cases include cost-efficient competitive programming, technical problem-solvi