Super User Daily: July 29, 2026

Opus 5 landed and the whole feed turned into a live benchmark, but the more telling move is where people are pointing it. Less "is it better at code" and more "read my Oura data, my four-thousand-note vault, my board's firmware, my stock research." The pattern holding across the day is that the harness keeps eating the model: weaker open weights jump a tier inside a good scaffold, and cheaper local rigs now run loops that used to need a data center. Underneath all of it the loudest thread is money and control, users watching Opus 5 drain a whole session in half an hour and asking why they can't just point the boring work at a cheaper model. OpenClaw, meanwhile, has quietly slid from "the future" to "is that still a thing."

People want to choose the model, not have it chosen for them. @Zoeillle wants to proxy OpenRouter and ChatGPT models inside Claude Code, and @svpino is fed up enough with Anthropic's stance on customers and open weights to jump to Codex.

Token burn is a black box and users want a flashlight. @lis186 describes context-window bloat pushing the model into a "dumb zone" of degraded reasoning, @Xudong07452910 flags prompt-cache expiry silently inflating bills during multi-agent idle waits, and @MAXIMISEART says the heavy MCP-and-subagent layers waste tokens versus a minimal harness.

The babysitting is wearing people down. @net_spheres calls granting permissions all day exhausting, @wolframs91 wants less constant hand-holding, and @TimSweeneyEpic watches Claude Code go on a markdown-file-creating rampage instead of solving the actual problem.

Context keeps dying at the seams between tools. @atulkumarzz and @heyDhavall both want one shared memory across agents and devices so switching tools doesn't mean re-explaining the whole project from day one.