Theo drops Opus 5.5 to high, swyx says go lower
Two shipping engineers now say the maxed reasoning dial is the wrong default, and neither posted an eval.
Theo said he moved Opus 5.5 from extra high to high a week or two ago after Julius convinced him to, and reports the lower setting is better almost all of the time. That is a reversal of his own configuration, not a benchmark result.
swyx went further in reply, pitching the other end of the dial. His argument is that low effort plus a critique loop beats what he called superultramaxhighwhatever, with the caveat that this is advice for the throughput-constrained, and he pointed at his own babysit-runs skill as the thing that makes it work.
low effort + critique loop > superultramaxhighwhatever
The Karpathy data point
swyx also noted that Karpathy is running medium on Opus 5.5 for his latest project, Hackertube, which launched that day. So among the three settings mentioned in the thread, extra high is the one nobody is defending.
having strong opinions on effort levels is like the tabs vs spaces of our time
None of them posted an eval. Theo's claim is a personal impression across his own work, swyx's is a workflow preference with a skill attached, and the Karpathy mention is an observation about someone else's setup. That is worth knowing before anyone reconfigures their harness on the strength of it.
Why the dial moved
The implicit claim in both posts is that reasoning effort trades latency and token spend against quality, and that past some point the extra thinking stops paying for itself. swyx's version adds a second mechanism. If you can run more passes for the same budget, a cheap first draft plus an explicit critique step can land in a better place than one expensive attempt, because the critique catches things the single pass would not.
swyx then undercut the whole argument himself, which is probably the most useful line in the thread for anyone deciding how much time to spend on this.
For working engineers, the practical read is that effort level is a per-task knob rather than something you set to maximum once and forget. The people posting here are tuning it by feel, and they say so.
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.