On crypto tasks, the cost-accuracy frontier is not what general benchmarks suggest. GPT-OSS-120B anchors the cheap end; the GPT-5 family holds the top. Claude Sonnet 4.5 and Gemini 2.5 Pro sit inside…

Fine-tuning does not close the gap. We trained LoRA on three base models directly on the full benchmark, with complete test exposure, and the learning curves stayed flat. The bottleneck is reasoning…