AI Lab · by Sanne de Vries · 2026-10-11
Claude Sonnet 4.5 is quietly becoming the default for agentic coding - anyone else noticing this?
Been running Claude Sonnet 4.5 through Claude Code for the last few weeks on a mid sized Go monorepo and the diff quality has genuinely surprised me. It plans multi file refactors, runs the test suite itself, and actually stops to ask when it's unsure instead of guessing. Compared to a year ago when every model would happily hallucinate an API, this feels like a real step change in reliability over raw cleverness. Curious what reasoning effort / thinking budget settings people are using day to day and whether anyone has pushed it on a genuinely huge codebase (500k+ LOC) without context rot setting in.
0 sparks · 11 comments
Comments
Imogen Clarke · 0 likes
same experience here. i run it on a rails app thats like 8 years old and it actually reads the surrounding tests before touching anything. didnt expect that from a model update tbh
Angel Reyes · 0 likes
yeah the test-first behavior is the biggest unlock for me, catches regressions before i even see them
Noé Petit · 0 likes
Honestly the reasoning budget stuff feels like a dial nobody explains well. I just leave it on high and eat the latency, worth it for anything non-trivial.
Lucas Williams · 0 likes
counterpoint: on our 1M+ line C++ codebase it still loses the plot after like 40 tool calls. great for focused tasks, not a magic bullet for huge legacy stuff yet
Mehdi Benali · 0 likes
this matches what i've seen too, context management is still the bottleneck not raw model smarts
Sofía Hernández · 0 likes
the thing that got me is it stopped guessing and just asked a clarifying question mid-task. first model that felt like it respected my time lol
Valeria López · 0 likes
cost is the real convo nobody wants to have. its great but burning through credits fast if you let it run wild on autopilot
Narin Chai · 0 likes
been doing side by side comparisons with gpt and the gap in agentic tool use is noticeable, claude just commits to a plan and executes cleaner
Midnight Runner · 0 likes
anyone tried pairing it with a strict linter/CI gate? curious if it self corrects faster when the feedback loop is tighter
Chiara Romano · 0 likes
yes! wiring it to fail fast on lint errors cut my review time almost in half
Maya Williams · 0 likes
500k LOC repo here, holds up fine as long as you point it at the right subdirectory instead of letting it roam the whole tree