@jeremyphoward@AnthropicAI Current API costs literally make anyone building their own agents impossible to be viable. I charge my app users $20, for them to burn this in <10 prompts, without earning a single penny from it.
LLM pricing is bonkers. Everyone is complaining about Opus using 90% of their 5h limit with one prompt alone. If Anthropic and OpenAI, which btw, heavily subsidise LLM costs, cant give anyone enough usage,
Then startups reselling tokens at $25/M tokens are cooked.
We need cheaper models immediately.
Day 9 of decoding startup trends.
The REAL reason Cursor is failing.
Over the past many months, many users quietly shifted to Claude Code or Codex. Including me!!
Why? Around January this year, Cursor suddenly became extremely expensive.
The problem is that LLMs are genuinely getting more and more expensive to run. To run Opus4.x @$5/$25 pricing, your $20 subscription would get over in 45mins?!
Anthropic and OpenAI force everyone to use their model at exorbitant API rates, and themselves serve users at pretty low rates (a GPU costs $2/hour).
Thats how OpenAI and Anthropic killed Cursor.
Own the infra if you want to win the AI race.
I spent the last few months using Claude Code CLI. Got pretty comfortable as well.
Until I was brutally reminded why CLI sucks in the first place. Cant scroll easily, cant copy, fixing typos is a nightmare.
Switching to GUI plugin and never going back, thankyou.
Your only job as a vibe coder is to provide your agents with a good feedback loop. Your agents will figure out the rest.
No exceptions, not even UI design.
Why does Codex start compacting the conversation at 220K tokens when the context window is 258K long?
Its because of the internal system prompt that is prepended on the cloud before running your request.