Claude by Anthropic
A family of foundational AI models
Start new thread

What's the best AI model for coding?

by•

New AI models pop up every week. Some developer tools like , , and let you choose between different models, while more opinionated products like and default to 1 model.

Curious what the community recommends for coding tasks? Any preferences?

3.6K views

Add a comment

Replies

Best

D) All of the above?

  • Write/negotiate product brief with ChatGPT 5.2 web thinking high.

  • Write/negotiate architecture+plan, based on the brief, with Opus 4.5 in Cursor

  • Implement the doc set (brief+architecture+plan) with Codex 5.2 CLI

  • Debug if necessary with Gemini 3 -> Sonnet 4.5 -> Opus 4.5, in Cursor, in that order, if the bug is being difficult.

  • Grok or Gemini 3 for codebase questions (where's this thing?)

Attempting to get some of the more complex thinking from Opus 4.5 without burning tokens implementing everything with it. I sometimes get through 3-5 briefs/features in a day by running in parallel, so my token burn gets pretty steep.

I find that enough documentation helps most models get decent results, but I do feel a difference with the frontier high thinking in Opus and Codex. Less to clean up when finalizing the feature. Fewer bad-coder behaviors like deleting tests that are failing to get a passing test suite.

Any open source model for coding?

 Devstral 2 (Small) by is released under Apache 2.0 afaik

thanks will give this a shot, also any good models you know for good frontend design results?

 good q! personally i had good frontend results with both Opus 4.5 and Gemini 3. hope it helps

 ICYMI Moonshot just announced Kimi K2.5, "the strongest open-source model to date for coding, with particularly strong capabilities in front-end development." () currently free on btw.

hope it helps!

Definitely Sonnet 4.5, with occasional Opus 4.5 mixed in when it can't handle the task. It's pretty crazy how quickly it's improving too.

Still significant hallucinations, but a good can dramatically reduce the ones that repeatedly pop up (e.g. assuming a certain testing framework, etc.)

a good can dramatically reduce [hallucinations] that repeatedly pop up

spot on - btw if you use , they recently included bundled docs for agents and it significantly improves their performance results

[1]:

My go-to is GPT-5.2

 do you use 's model for every type of task (plan, code, debug)?

Claude Sonnet 4.5 for me.

Consistent, predictable, and easier to work with over longer sessions.

is leading the way!

Used all except devstral. Sonnet 4.5 gives the best "to the point" ability for larger projects without overdoing or derail.

Opus 4.5 missing from this list, isn't it? ;)

problem for most of us: who has the budget to pay for most and compare? yikes

How can a single developer know what actually works across different setups and what's just marketing hype? especially on "hope it is free" budget...

Most of us do NOT have unlimited personal budgets so, I for one, have relied on free tools that , in exchange, I contribute to train with a thumbs up or a thumbs down.

My current setup is VSCode running Windsurf, the default model, which is free. I also have a local version of deepseek running on my Mac. Windsurf has saved me hours it is very good at quick autocomplete without trying to write everything itself. it has also helped me refine existing code to be shorter and more readable. it's helped me find places where I am not error checking as well.

Deepseek I use mostly to explain code but I find it to be subpar as far as generating good code that is usable. However it also took me a long time and lots of reading to settle on Windsurf in particular. It was overwhelming, the number of choices.

I have started narrow basically focusing on refactoring code to be more efficient or autocomplete for long JSON files. However I am seeing that I can probably expand to generating documentation.

What are you using in your own development environment and what do you see as the pros and cons? I am seeing that cost is definitely a limiter in my case, so I settled on a free tool that also limits the number of models it uses.

 Totally relatable. I spent a long time on free tiers myself until I realized I was spending more time "making it work" than actually saving money. Here’s what helped: many tools (including Cursor) offer trial periods or $5–10 credits - you can run your real tasks through different models over a week and see what actually speeds things up. Then the math’s simple: if a model saves you 2 hours a week, it pays for itself even at $20/month

Windsurf is a solid pick too, their request caching actually works and doesn’t eat tokens on repeated edits

Cursor or claude with blocpad

I use Antigravity from google, and mainly use Sonnet and Opus. Its weird but on the same prompt and context, sometimes, Gemini models hallucinates and start rewriting the whole files

First
Previous
123
•••
Next