AI Lab · by Jiho Lee · 2026-10-10
Claude Sonnet vs Opus vs GPT-5-class models for actual daily coding — what's your real setup?
Past few months I've bounced between Claude Sonnet for fast iterative coding, Opus when I need it to actually reason through a gnarly architecture decision, and GPT-class models for certain refactors. Reasoning budget/thinking tokens make a bigger difference than raw model size in my experience, especially for debugging. What's everyone's actual day-to-day stack look like right now, and where do you still feel the gap with agentic coding tools (Claude Code, Copilot, Cursor, Windsurf)?
1 sparks · 9 comments
Comments
Khalid Al Zahrani · 0 likes
Sonnet for 80% of my work, its just fast enough that i dont lose flow. I only reach for opus when sonnet gets stuck in a loop on the same bug twice
Hamza Malik · 0 likes
same experience here, the 'stuck twice = escalate' heuristic is basically my whole workflow now
Salma Idrissi · 0 likes
Honestly the agentic tool matters more than the underlying model at this point imo. same model different harness gives wildly different results
Khalid Al Zahrani · 0 likes
running local qwen or deepseek distills for anything thats privacy sensitive, its not as good but good enough for 60% of tasks and free after the hardware cost
Lotte Jansen · 0 likes
whats your setup, ollama on an m-series mac?
Haruto Sato · 0 likes
the thinking/reasoning budget knob is honestly underrated, i wish more products exposed it instead of hiding it behind a 'mode' toggle
Maya Williams · 0 likes
unpopular opinion but i still hand write a lot of boilerplate bc reviewing ai code carefully takes almost as long as writing it for me
Lucas Williams · 0 likes
cost adds up fast if youre not careful, opus on a big repo context burns through credits stupid quick
Salma Idrissi · 0 likes
my team switched entirely to agent based workflows this year, feels like a totally different job than 2023 tbh, for better and worse