AI Lab · by Mina Chowdhury · 2026-10-11
Claude Opus 4.6 vs Grok 4.6/4.7 for real agentic coding — my 2 week head to head
Ran the same gnarly refactor tasks (100k+ LOC monorepo, legacy Java + a Next.js frontend) through Claude Opus 4.6 and xAI's Grok 4.6 then 4.7 inside Cursor. Opus is still the most 'careful' — fewer silent assumptions, better at asking before deleting stuff. Grok 4.7 genuinely surprised me though, it's noticeably cheaper per token and got closer to Opus quality than I expected on multi-file edits. Reasoning budget settings matter a LOT, default low settings make every model look dumb. Anyone else benchmarking these against real repos instead of toy leaderboards?
0 sparks · 10 comments
Comments
Ananya Rao · 0 likes
the reasoning budget point cannot be overstated, i wasted a week thinking a model was bad before realizing i had it on the cheapest tier lol
Daniela Torres · 0 likes
same mistake, cost me actual sleep debugging 'model hallucinations' that were just truncated thinking
Brian Otieno · 0 likes
we run both in prod for different tasks. opus for anything touching billing/auth code, grok for bulk mechanical refactors where speed matters more than paranoia
Ren Suzuki · 0 likes
grok 4.7 has been a genuine shock to me too, i was fully ready to dismiss it as hype given xAI's history but the agentic tool-calling reliability is there now
Rafael Almeida · 0 likes
no mention of cost per successful task though which is the metric that actually matters for a team budget, not just raw quality
Noah Johansson · 0 likes
fair, roughly for us opus comes to about 2.5x grok's cost for maybe 15% fewer fixups needed. your mileage will vary wildly by codebase
Sophie Gagnon · 0 likes
counterpoint: toy leaderboards exist because real-repo benchmarking is incredibly noisy and non reproducible, which is exactly what's happening in this thread lol, no offense
Quiet Runner · 0 likes
java + nextjs monorepo sounds like my personal nightmare, respect for even attempting automated refactors there
Solar Mango · 0 likes
anyone tried qwen3 or deepseek v4 for this same task? curious how far behind the open models actually are now
Curious Baker · 0 likes
the 'asks before deleting stuff' behavior from opus has saved my bacon at least twice this month, worth the premium alone imo