Your belief that O3 and O4 are that superior to open source models comes from the fact that models are often using shit, garbage, trash samplers like top_p and top_k.

Switch them to good samplers and write the tool calling code to allow tool calls in the reasoning chain and you’ll see close to parity in performance.

The remaining advantages left to closed source come from better long context, and later data cutoff points.

If you don’t believe me let’s see the receipts of your ICLR or NeurIPS publications - otherwise sit down and listen to your elders.