Ask HN: $50 monthly budget, which coding models would you recommend now?
I currently have a claude pro monthly subscription ($20) which I use for coding. It's been useful but I'm fatigued from optimising my work around it's session limits. There are so many choices and providers out there today but hard to get a good signal about what's good. I'm not looking for another Opus-level model but something reliable enough that it can follow TDD well.
You can swicth to Sonnet 4.6. It is even an fantastic model for coding, without changing you subscription.
Here the 5 hours window limit is much more larger, if using it with specific tasks.
If you want to change the LLM, my advice is Kimi K2 that's much more chip and very skilled for coding.
(you are right: Opus is a very token-consuming even in my experience, but often it is over-skilled and you don't need it)
Hey man, here's how to properly spend your 50 bucks and get the most bang for the $$:
opencode go - $10 a month gets you decent allowance with latest GLM, MiniMax, and Kimi.
Windsurf - $20 gives u access to nearlly all models.
For me, I'm having these two subscriptions plus: shared Codex plans & shared Kimi coding Allegretto from Chinese resellers for around $10.
I only use the Claude Opus/Sonnet inside Windsurf after everything is crystal clear, or I want to let it handle tricky stuff where other models can't address. I use GLM/Kimi for everything else. And Windsurf offers free models all the time.
Codex? I know OpenAI is really politically unpopular right now, but it has very high usage limits for the $20 plan. Claude ($20) and Codex ($20) are hard to beat in terms of pure value. Just set Codex on Thinking-High/Extra-High, and it is Opus 4.6 levels for sure (although both have their niche, where they're superior).
Claude for the default 5-hour limits. Cursor for the monthly limits when those run out. Cursor is one of the more efficient ones for dumb uses... if you're typing out your own tests, it should work great.
A lot of people recommend OpenAI Codex lately, but I feel it's more suited to vibe coding, where you're giving it a very high level idea.
Given the hard budget, I'd throw the $30 on API credits and use those as overflow. If you can do the $100 plan though, that would be optimal. It's a frustrating game though. I'd love to see a 50 dollar plan. At least Claude has a 100 plan, unlike openai.
I did try Extra Usage and API credits with Claude and it chews through $5 in minutes. Unlike GPT, the subscription is a better deal. I definitely do not recommend API credits and such. It's more efficient to use Claude through some other provider like Antigravity, Cursor, and so on.
Whoa, if you're using it that much, you should definitely look at the $100 plan. Either that, or you'll have to use an older model or reduce thinking (which does really tank the quality of results unfortunately) or look at a local model. I don't think there's really many other options.
For a while I did a $20 sub to both OpenAI and Claude and just switched back and forth. I was able to get pretty far that way, but it can feel a bit disruptive.
If you want to change the LLM, my advice is Kimi K2 that's much more chip and very skilled for coding.
(you are right: Opus is a very token-consuming even in my experience, but often it is over-skilled and you don't need it)
opencode go - $10 a month gets you decent allowance with latest GLM, MiniMax, and Kimi. Windsurf - $20 gives u access to nearlly all models.
For me, I'm having these two subscriptions plus: shared Codex plans & shared Kimi coding Allegretto from Chinese resellers for around $10.
I only use the Claude Opus/Sonnet inside Windsurf after everything is crystal clear, or I want to let it handle tricky stuff where other models can't address. I use GLM/Kimi for everything else. And Windsurf offers free models all the time.
A lot of people recommend OpenAI Codex lately, but I feel it's more suited to vibe coding, where you're giving it a very high level idea.
Whoa, if you're using it that much, you should definitely look at the $100 plan. Either that, or you'll have to use an older model or reduce thinking (which does really tank the quality of results unfortunately) or look at a local model. I don't think there's really many other options.
For a while I did a $20 sub to both OpenAI and Claude and just switched back and forth. I was able to get pretty far that way, but it can feel a bit disruptive.