I look at the output of Kimi and the costs of running inference on it that i can replicate, and it isn't that bad, although admittedly i don't have to worry anywhere near as much about scaling it and about having to dedicate large amounts of compute to research and distillation on the back end. It's true that it's perhaps a step behind SotA vs January's Opus or current Codex, depending on what you do. But not by a lot. In fact it's leaps and bounds superior to the current subscription API experience. Together with GLM, Qwen and Minimax they are an amazing backstop just the way they are right now.

With all the layers of obfuscation it's hard to even know roughly how many i/o Opus tokens do Claude subscriptions pay for. They'll give you some flippant arguments like "people were not looking at thinking so we're not showing you anymore" with a straight face. However podcasts still insist Anthropic are "winning the AI war" (??) it really makes me wonder because in no metric I can see them as providing neither best value nor best quality, and let's not get started about consumer experience.

My intuition is that things must be really bad so they're willing to pull the kind of moves they're pulling right now. They're speedrunning people into understanding how important it is to be able to run your own generative AI infrastructure for reliability, thus becoming a very fancy but trustless throwaway solution factory.

I wonder if OpenAI will turn the screws similarly if/when their pockets start to dry up at a certain pace.