Can’t read the whole article bc paywall… But if they are really worried about token cost for converting PDF to a Powerpoint, they ain’t seen nothing yet. Agentic coding the way AI companies push it (multiple agents in parallel with Claude, looping etc) uses way way wayyyyy more tokens than this.
How do you use it without agents? Just prompt the chat interface?
When I used claude code, I’d go through the smallest tier subscriptions 5h limit in half an hour sometimes if I used Opus lol, reckon at API pricing it would’ve been 100 bucks an hour
Chat prompts and then accept the changes when they are worth using. LLMs are entropy apparatuses, injecting that shit straight into your code base is crazy. It took one module/subsystem done with agents for me to realize that it was completely unmaintainable. Not a single eng can make a change to it and the only option is to rewrite. So I banned agents. I know some people still try to get away with that style of LLM coding but I encourage people to instant deny a PR if it looks like garbage.
Agent mode is disabled, only chat mode works. This is with GitHub Copilot harness in JetBrains Rider or Visual Studio Pro. Agentic coding means it can make decisions without a human.
Can’t read the whole article bc paywall… But if they are really worried about token cost for converting PDF to a Powerpoint, they ain’t seen nothing yet. Agentic coding the way AI companies push it (multiple agents in parallel with Claude, looping etc) uses way way wayyyyy more tokens than this.
Archive link: https://archive.ph/c6DSo
Have you used Opus 4.8 at API costs? Without using agents, I can burn $20 an hour no problem. I use Kimi K2.6 and GLM 5.2 these days.
How do you use it without agents? Just prompt the chat interface?
When I used claude code, I’d go through the smallest tier subscriptions 5h limit in half an hour sometimes if I used Opus lol, reckon at API pricing it would’ve been 100 bucks an hour
Chat prompts and then accept the changes when they are worth using. LLMs are entropy apparatuses, injecting that shit straight into your code base is crazy. It took one module/subsystem done with agents for me to realize that it was completely unmaintainable. Not a single eng can make a change to it and the only option is to rewrite. So I banned agents. I know some people still try to get away with that style of LLM coding but I encourage people to instant deny a PR if it looks like garbage.
If you accept changes then you’re probably using an agent though?
Agent mode is disabled, only chat mode works. This is with GitHub Copilot harness in JetBrains Rider or Visual Studio Pro. Agentic coding means it can make decisions without a human.