Corrections

When we get a claim, a number or an attribution wrong, we correct it on the original post and in the next issue, and we list it here. This page is the record.

A typo, a broken link or a rewritten headline is not a correction. We fix those without a note. A correction is anything a reader could have acted on that turned out to be wrong.

If you think something we published is wrong, reply to any issue. Include the source that says otherwise.

On the newsletter

• 2026-09-23 · Claude Opus 5.5 and GPT-6 Sol ship hours apart (Sep 23)
  What we said: We set Anthropic's claim, which was about default settings, against a result Artificial Analysis measured at maximum effort.
  What's right: At default settings, Artificial Analysis measures Opus 5.5 at 37% of Opus 5's cost per task.
  Where the fix ran: A note on the original post (23:20 UTC), and the next issue
• 2026-09-04 · OpenAI ships GPT-6 Astra (Sep 4), item 6
  What we said: We wrote that the terms clause "names the two product accounts, not the Google account", as if Gergely Orosz had misread it.
  What's right: He had read the terms as they stood.
  Where the fix ran: A note on the original post (16:44 UTC)
• 2026-08-13 · Zed ships Delta (Aug 13), item 3
  What we said: We said the fuel for the story was a viral thread of complaints.
  What's right: The fuel was Anthropic's disclosure that Monday that new Claude models embed text watermarks.
  Where the fix ran: The original post, edited the same morning; the note was added on 2026-10-01
• 2026-08-06 · Meta ships Muse Code (Aug 6), the TCRF item
  What we said: We said there was no statement from TCRF.
  What's right: TCRF's co-founder had been responding publicly on Bluesky for hours before the issue went out. The item now quotes and links it.
  Where the fix ran: A note on the original post, the same day
• 2026-08-04 · An npm worm is loose (Aug 4), the lead
  What we said: We ran Aikido's first count, 868 compromised packages.
  What's right: Aikido revised the count to 434 after publication.
  Where the fix ran: The original post, edited the same day (the note was added on 2026-10-01), and the next issue

On our social posts

• 2026-09-29 · Bluesky reply (Aug 6)
  What we said: We said Zenity's report found that more than 30% of dangerous skills abuse Claude Code or OpenClaw as malware droppers, and that one reinstalls itself if deleted.
  What's right: Neither claim is in the report. Both came from Zenity's press release.
  Where the fix ran: A reply under the original post, 54 days later
• 2026-09-23 · X and Bluesky post on Opus 5.5 pricing (Sep 22)
  What we said: We implied Anthropic's 40% saving was in doubt, citing results measured at maximum effort.
  What's right: At default settings, Artificial Analysis measures Opus 5.5 at $1.34 a task against $3.61 for Opus 5, 63% less.
  Where the fix ran: A reply under the original post on X and Bluesky, and the next issue