The industry seems to be discovering that good AI is really expensive to run:
- Anthropic did a “test” in their sign up flow removing Claude Code from their $20 subscription, then added back after a community backlash, without much communication
- Github Copilot, one of the best deals out there, removed the most expensive models from their offering, added session and weekly limits, and blocked new sign ups for all their plans
Its clear that those coding agents consume a lot of tokens, and the AI providers were simply not ready for the upcoming demand. Most of the AI performance increase we got in the last month is a combination of more tokens and bigger models, basically throwing compute at the problem.
A few thoughts out of this:
- OpenAI didn’t made any changes yet, which can suggest they have more compute and cash available to burn. Codex might see increased usage being the better deal
- Open Weights AI is even more important. Luckily we have GLM and Kimi around that offers near frontier level of performance for a fraction of the cost