What models have been successful for you?

[clsferguson]

Most seem to work for everything, except coding.
sonnet 4.5 has been the best for me so far for coding. but it has an issue with bug fixing. It rushes to fix the first thing it sees and thinks it's fixed.

I am trying to use cheaper models.

I tried grok-code-fast-1 -> unable to implement. Not enough context window is my guess.
gpt-5-codex -> actually goes in slow circles and gets nothing done.

I end up getting frustrated and just shelling out for the extra tokens

[wilsonneto-swe]

For me Sonnet 4.5 is the one to go, so far it is going pretty well.
Most of my problems so far was regarding too much context, too big features and so on, not really regarding the model.

[pkoetzing]

GPT-4.1 works pretty okay for me. I use Sonnet 4.5 only for spec and implementation. Grok is bad.

[lynglink]

Would a stage specific model approach be suitable? Based on best practice use of spec kit workflow?

[roee-hersh]

my conclusion for using speckit on daily basis for 3 weeks now,
this is my experience working with almost all the models (on cursor and copilot)

Planning

1. GPT 4.1 (when its not down)
2. Gemini 3 Pro is pretty good and cheaper alternative
3. Composer 1 was surpassingly good for planning

i od not recommend planning with sonnet 4.5 - his trying too hard which causes him to overflow the MD files with content which make the context explode while already costing 30% more

Coding

1. Composer 1 - this is my winner with a good spec and plan this guy wins every time not despite but because his stupid and blindly follow the tasks like any execution model should
2. GPT Codex - smarter alternative pretty good but slow
3. Sonnet 4.5 - very expensive but efficient, he doesnt always follow the tasks and generate "side quests" for him self (like creating summary.md files), his not good at not doing stuff, for example try to ask him to implement only the first task or "phase 1 only", most of the times he will just do everything - but overall his pretty good