In your opinion, what is the most economical yet reliable AI model for everyday code development, i.e., for tasks that are not too difficult?
I look forward to hearing your opinions based on your experiences.
1 Like
GPT models for sure until sonnet 5 comes out
2 Likes
Cost-wise for me:
For coding stuff - Grok code fast 1 (was free, but now paid, but still the cheapest option)
For asking questions about the codebase - Gemini 2.5 Flash (yes its still a great cheap option available)
For planning - No cheap model here, but Opus 4.5 thinking and GPT-5.2-xhigh are great candidates. Although you can try GPT 5.2 or Gemini 3 flash also, they are okayish and cheaper than the big boys.
For more model pricing check here: Models | Cursor Docs (sort the output column ascending)
3 Likes
Grok Code does the job and is dirt-cheap.
It still requires well-defined tasks that do not contain too many steps. But for any “implement UI component” or “create ansible role” it is all you need. The same is true for following detailed plans or cloning existing patterns.
1 Like
Unfortunately, I’ve noticed that Grok Code makes beginner mistakes even on simple code. I think I’ll stick with Sonnet 4.5, which is quite expensive but at least doesn’t make any major mistakes.
1 Like
For simple code, you might want to give gpt-5.1-codex-mini-high a shot. It’s scores better than grok-code by a mile and even beats the non-reasoning sonnet 4.5. Plus its cheap!
2 Likes
GPT-5.3-Codex
1 Like
GPT-5-mini is super cheap or free and for everyday tasks that have small contexts it is very effective.
1 Like
Grok code is like a kid with ADHD running out the door on the way to school gets halfway there and forgot his school bag.
2 Likes
Kimi K2.5
Fast and reliable for speed. Since it’s hosted on Fireworks, the performance is really good. I don’t think it reaches 200 TPS, but I believe it’s at least around 100 TPS.
Hi Naufaldi,
how does Kimi perform in software development? Does it create good code?
Bye
Yes, it works great, often on par with Codex 5.2 Medium. I don’t see any major issues, especially for frontend tasks.
There are some cases where Kimi can’t fix a problem. When that happens, I usually switch to Codex 5.2 Medium, and Codex is able to solve it. But most of the time, I use Kimi as my daily driver and rarely run into problems.
That’s why, for my daily 9–5 engineering work, Kimi is my go-to. My usage shows it too: around 157M tokens for Kimi, compared to 52M for Codex.
1 Like
Thanks for your reply, Naufaldi. I’ll try that too.
bye