The most cost-effective way I found to use their models is to use them through a Github copilot licence. Github charges by request, not per token. Asking Opus with high effort to plan a feature in depth "costs" the same as asking it the number of "r" in strawberry.

I've got a setup where GPT5-mini (Free on GH) talks to you to refine and draw the outline of your feature, calls a single Opus subagent to plan it in depth, then calls a single sonnet subagent to implement it.

Github will only charge you for one opus request (With a 3x multiplier) and one sonnet request, whether they consume 50 or 500.000 tokens. I'm running this setup for 9 hours a day at work and I've barely consumed 40% of my monthly allowance.