/AI1d ago

Researcher Flags Unexplained Spike in Claude Opus Token Use After Launch

--0--
Original posts
Christopher Potts#222
Quote posts
Christopher Potts#222
Original post
Christopher Potts@ChrisGPotts#222inAI

Does anyone know why Opus 4.6's token consumption in Claude Code would skyrocket specifically in the period Feb 20 to Mar 4, 2026? The model launched Feb 5 with default reasoning "high", but the steep increase for this model doesn't happen until ~2 weeks after launch.

Also noteworthy: Mar 4 is when Anthropic changed the default reasoning to "medium", but this seems only to slow the rising consumption rather than reversing it.

These observations are based on SWE-chat. We (@mmooritz and I) can control for lots of factors in this dataset (user, project, session length, etc.), and we can't find an explanation for the increase. My own hunch is that there was an unreported change to the reasoning around Feb 20, but I could be missing something.

1:23 PM · May 31, 2026 · 6.2K Views
Sentiment
Sentiment unavailable for this story.
Cluster Engagement
-
Views
-
Comments
-
Reposts
-
Bookmarks
Expand data
Posts from X
Most Activity
Most ActivityTimeline
VIEWS4.8KBOOKMARKS7LIKES20RETWEETS4REPLIES4
Christopher Potts@ChrisGPotts

Here's a plot that summarizes this puzzling situation. We can't figure out the cause of the rise in Opus 4.6 token usage around Feb 20. The Mar 4 high-to-medium change only slows the trend too. (The "persistent cohort" helps us feel confident that this is truly an overall trend.)

Christopher Potts@ChrisGPotts

Does anyone know why Opus 4.6's token consumption in Claude Code would skyrocket specifically in the period Feb 20 to Mar 4, 2026? The model launched Feb 5 with default reasoning "high", but the steep increase for this model doesn't happen until ~2 weeks after launch.

Also noteworthy: Mar 4 is when Anthropic changed the default reasoning to "medium", but this seems only to slow the rising consumption rather than reversing it.

These observations are based on SWE-chat. We (@mmooritz and I) can control for lots of factors in this dataset (user, project, session length, etc.), and we can't find an explanation for the increase. My own hunch is that there was an unreported change to the reasoning around Feb 20, but I could be missing something.

9hViews 4.8KLikes 20Bookmarks 7