$200 per month can be a lot of money, especially outside of America (which enjoys an absurd median income of $45,140, vs global ~$3,480). So I think for the egalitarian option, it's best to look at what the more affordable $10/month OpenCode Go subscription (still expensive in many parts of the world) gets someone - roughly 37,800 Deepseek V4 Flash requests a month! And if one is in a supported region, 226,600 Muse Spark 1.3 requests a month (at the cost of having their data trained on)! Muse Spark 1.3 gets 24.9% on CritPt (good physics benchmark, proxy for very hard mathematical/scientific reasoning ability), and Deepseek V4 Flash 0731 gets 16.6%. GPT-5.3 Codex scores 16.9% on this benchmark, and GPT-5.4 23.4%. So OpenCode go offers access to approximately the public frontier from ~6 months ago. And the frontier 6 months ago was already used to do meaningful math and theorem proving! So I think overall LLMs that are good at math are pretty globally accessible, but Astra/Fable-level models most definitely aren't. Perhaps another 6 months and they will be!