i tested sonnet 4.6 in openclaw for a day and it meaningfully improved the agents. here’s a brain dump of handful of things i noticed: the biggest thing is it follows instructions precisely, and all instructions. seems like it gobbles up every markdown file in the workspace before acting. i asked for new formatting and it responded “i checked memory and didn’t find your preference…” opus never did that, it cherry picks what to take in as context before doing something. and suddenly there were a handful of cron jobs it started updating me on that i didn’t know existed. opus was running them quietly in the background even though i asked repeatedly not to work silently in the background. it’s more empathetic. idk, this one is visceral, i can’t put my finger on it. but the way it responds is less sycophantic. it will admit when it’s not sure about something. which makes me want to have real dialogue with it instead of the usual one or two word prompts i use to nudge it to do something. and i think this is in part why it’s a better writing assistant. any drafts it creates for new content is actually usable. whereas any content creation help opus and other models has been dead on arrival, no matter how hard i argue with them. the self-reflection loops seem to be more effective. two examples: 1. “write content then after i post use the browser to track analytics, think about what worked/didn’t work, and apply your learnings in the next draft.” for the first time, it did. 2. “book me a table at [hard to get restaurant]. update your approach after each failed attempt.” for the first time, it stopped polling for cancellations and researched when new tables dropped on its own. i also have a food log, all of my workouts, and a dexa scale that shoots my bmi/body fat to a webhook. opus rarely looked at all of them before recommending meals or workout adjustments. sonnet referenced all 3 any time it proactively pinged me about fitness. “cut your fruit intake and get more starchy carbs in today because you have a big workout coming up” or “try to drink more water today because the packaged meal lunch you had earlier has a lot of sodium” ultimately i think it just comes down to the bigger context window and more emphasis on following instruction. it might not be the best general model, but it feels like it was tailor made for openclaw.

just got this message. first time it actually feels like a competent health/fitness coach