It would be very useful for companies to isolate interesting programming challenges in their past and publish evals on them (without revealing the actual codebase). In theory companies adopting these models should already be doing this to evaluate cost/benefit for each model, so it would be a matter of publishing them on a regular basis.