141–150 of 245 posts
Considering how much progress I made vs how much I paid, I couldn't make a scientific assessement, but it felt pretty close.
The only way AI will be profitable to companies like Anthropic or OpenAI is to make the cost $1000-2000/month or more for coding. Every programmer will be forced to pay for it because it's only a fraction of their salary (in the US anyway) and it's the only way the programmer will be competitive. Whether the company pays for it, or they pay for it themselves, it will need to be paid. There's no other way that these c…
I routinely match or beat Claude with regards to speed, I often race it to the solution because Claude just takes so long to produce a usable result.
Staying competitive doesn't mean only paying an AI for slop that often takes longer to produce. AI is a convenience, it is not the only way to produce code or even the most cost effective or fastest way. AI code also comes with more risk, and more cognitive load if you actually read and understand everything it wrote. And if you don't then you're a bit foolish to trust it blindly. Many developers are waking up to the reality of using AI, and it's not really living up to the hype.
Earlier quoted context omitted.
This feels a lot like the same playbook we’re seeing with dynamic pricing in retail, just applied to compute instead of products. You never really know what you’re getting, and the rules shift under you. What makes it worse is the lack of transparency. If there were clear, hard limits, people could plan around it. Instead it’s this moving target that makes it impossible to trust for real work. At some point it stops…
What a horrid glimpse in the future. I hope we won't get there and we all collectively fight back with our wallets.
I cancelled my pro plan last month. I was using Claude as my daily driver. In fact had the API plan also and topped it with $20 more. So it was around $40 each month. Starting from December last year it has been like this. When sessions could last a couple of hours with some deep boilerplate and db queries etc. to architecture discussion and tool selection. Slowly the last two months it just gets over. One prompt and…
Earlier quoted context omitted.
Once you get used to using claude as an abstraction layer you start getting pretty reckless with it. My organization has the concept of "premium models" where our limits reset every month. I hit my limit pretty quickly last month because I was burning tokens doing things that would have been a simple bash loop in the past - all because I was used to interfacing with Claude at the chat layer for all my automation need…
This is a real danger that I think a lot of people will run into as prices go up more and more in the future. Completely outside of the productivity debate, offloading cognitive tasks to LLMs leaves you less practiced in them and less ready to do them when the LLM isn't available. When you have to delegate only certain tasks to the LLM for financial reasons, you may find yourself very frustrated.
I find Claude code to be a token hog. No matter how confidently the papers say context rot is not an issue I find curating context to be highly important to output quality. Manually managing this in the Claude Webui has helped with my use cases more than freely tossing Claude code at it. Likely I am using both "wrong" but the way I use it is easier for me to reason about and minimize context rot.
I have seen more than one comment on this thread mentioning kimi though - I'll have to test it out.
qwen3-coder-next has been surprisingly capable as a local model too - needs to be used to make small changes where you know exactly what the final code should look like rather than implementing whole features, but it is free (except for the power bill).