Opus 5.5 fails twice as many T3 turns on max as on high
Telemetry from real T3 Code sessions, not a benchmark, and the extra failures cluster in the first 30 seconds.
Adam Holter reports that Opus 5.5 fails twice as many T3 turns on max reasoning as on high. The numbers come out of T3 Code's anonymized telemetry, which Theo says he handed to Holter specifically so more posts like this would get written.
max fails more in every turn-length bucket, and most of the extra failures happen in the first 30 seconds.
The obvious explanation is length. Turns on max take longer, longer turns have more chances to go wrong, and the failure rate follows. Holter checked and says that is not it. Max fails more in every turn-length bucket, and most of the extra failures land in the first 30 seconds of a turn, before a long run has had time to become long.
The pattern is not unique to Opus 5.5 either. Holter says Sol shows the same shape at a smaller gap, about 1.4x. He does not offer a cause, and his post ends with a shrug emoticon rather than a theory.
Why this one is worth more than the usual vibe
Most of the advice to turn the reasoning dial down has been anecdotal. Theo had already moved to high effort on Opus 5.5 while swyx argued for going lower still, and that exchange was two experienced people describing what felt better in their own work. This is aggregate behavior from real sessions in a shipping coding tool, which is a different kind of evidence. It is also not a benchmark, so it carries the usual caveat that attaches to production data. Whatever selects which turns get run on max in the first place is also baked into the comparison, and nothing in Holter's post controls for that.
Theo's reading is blunt. He framed the chart as "more reasons to never use 'max' reasoning" when he shared it.
For working engineers, the useful part is the 30 second detail rather than the headline ratio. If the extra failures were spread evenly across turn length, you would suspect the model wandering off on long tasks, the thing everyone already assumes about heavy reasoning. Failures concentrated at the start of a turn point somewhere else, toward how the turn is set up rather than how long it runs. Nobody in the sources knows which, including the person holding the data.
more reasons to never use "max" reasoning
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.