So I saw a person on X (via Google Discover) who did a much more thorough, downright impressive analysis on the value of AI subscriptions: going model by model, calculating the “effective per token costs”, and putting it all into a table with a new Pareto frontier derived from that.

The standout for me was Luna being practically cheaper than DeepSeek when accessed via the subscription! That’s a pretty high bar to clear especially compared to the prices before the raises.

Oh right, the two links below lead to X - a site with lesser privacy guarantees than I usually link to, and certainly less than this one. Please be mindful!

Chinese post (the one I saw first)

English post (found by ChatGPT)

But they went even further, and open sourced an entire GitHub repo with the data and methodology. So… go check them out! It’s actually extremely impressive…