Claude Sonnet vs Opus: when is the extra reasoning budget actually worth it?

Mara Fischer · 0 likes

we swapped our whole pipeline to sonnet w/ high budget for anything under 2k loc, opus only for the big architecture reviews. saved like 40% on api cost and barely notice quality drop honestly

Gentle Baker · 0 likes

same experience here, opus is basically our 'senior engineer escalation' tier now lol

Urban Fern · 0 likes

depends heavily on the task imo. for pure code gen sonnet is fine but for debugging gnarly concurrency bugs opus finds the root cause way more often in my testing

Mina Chowdhury · 0 likes

honeslty i think ppl overrate the reasoning budget knob, ive tested it on same prompts and the variance between runs is bigger than the variance from budget size

Nusrat Akter · 0 likes

this, temperature + sampling noise swamps the signal unless ur averaging over many runs

Mara Fischer · 0 likes

claude code specifically seems tuned really well for sonnet, i wonder if opus is even worth using there vs just raw api for architecture stuff

Rohan Mehta · 0 likes

anyone tried measuring tokens-to-first-correct-solution instead of just cost? thats the metric that actually matters for me

Angel Reyes · 0 likes

we do opus for planning phase then sonnet for execution. best of both worlds and cheaper than full opus end to end

Priya Nair · 0 likes

lol opus still randomly decides to rewrite my entire file when i ask for one small change. anyone else get that

Mango River · 0 likes

yes!! its so confident about it too, like it genuinely thinks its helping

Hugo Martin · 0 likes

not enough ppl talk about latency difference, opus thinking time adds up fast when ur doing agentic loops with 50+ tool calls