Opus 5.5 on max fails twice as often as high
Adam Holter went through T3 Code's anonymized telemetry and found Opus 5.5 fails twice as many turns on max reasoning as on high, with most of the extra failures landing in the first 30 seconds. Nobody has an explanation yet.
The big disagreement
Ball says the problems of remote coding are solved because you are not coding
Responding to a podcast episode with Armin Ronacher that asked whether local desktop software is dead and debated his own agent workflow, Thorsten Ball said the problems of remote coding are already solved.
For one, you're no longer coding :) No need anymore to get LSPs or other editor-centric tooling running. Or an editor!
Receipts
Holter: Opus 5.5 fails twice as many T3 turns on max as on high
Adam Holter, given access to T3 Code's anonymized telemetry by Theo, reports Opus 5.5 failing twice as many turns on max reasoning as on high. His first guess was that max turns simply run longer, but max fails more in every turn-length bucket and most of the extra failures happen in the first 30 seconds.
more reasons to never use "max" reasoning
How they actually work
dax put every subagent on DeepSeek Flash and kept Opus as the orchestrator
dax says his team spent the week switching all subagents to DeepSeek Flash with Opus 5.5 only on the primary agent, and that it is working great this time where an earlier attempt was not.
ask for proof of work instead of a summary