← The AI Hype Audit — all 550 verdicts
PARTLY
Partly true: Opus 5 announcing work and then ending its turn without doing it is a documented, reproducible-enough bug report, but 'every day, multiple times a day' and 'ignores check-ins' are one team's experience, not an established pattern
The claim'New to Opus 5: it accepts work, lists the exact steps it will take, and then does nothing. No work actually happens. When I come back an hour later and ask for the status, it doesn't admit it did no work; it silently ignores me and starts working. This has happened every day I've used Opus 5, multiple times a day.' 'Anthropic, you've got a lot of work to do before Opus 5 can handle heavy workloads from power users.'
The core behavior is real and has been reported independently. A public GitHub issue on Anthropic's Claude Code repo (#83713, opened Aug 4, 2026) describes exactly this: claude-opus-5 ended its turn with a present-tense 'updating the note now' message and no tool call, so nothing happened, and the user only noticed minutes later. That reporter's own logs showed it was rare (2 confirmed cases out of 23 risky turns) and did not occur with older models in about 640 comparable turns. The issue got no reply from Anthropic and was auto-closed as stale on Sep 6. Separately, an Anthropic Claude Code engineer publicly acknowledged Opus 5's output was 'spiky' and said the team was working on consistency, while saying internal evals showed no regression. Anthropic's status history also shows several Opus 5 error incidents in July to September 2026. What does not hold up as a general claim is the frequency. 'Every day, multiple times a day' is one team's experience with their own setup and prompts, not something the public record confirms. 'Ignores my check-in and just starts working' is, read plainly, the model resuming the task when prompted; it should also say it had not started. Opus 5 is now listed as a legacy model, with Opus 5.5 current since late September. The poster is selling a members-only livestream.
What holds up
- Reel transcript and on-screen cards read: 'It accepts, it announces, it does nothing'; 'every day, multiple times a day'.
- GitHub anthropics/claude-code #83713 (Aug 4, 2026): Opus 5 ends turn on 'doing it now' narration with zero tool calls; 2 phantom cases in 23 at-risk turns; none in about 640 turns on older models; closed as stale Sep 6 with no maintainer reply.
- Anthropic engineer (via Digg/X coverage) called Opus 5 'spiky', said work on consistency was underway and internal evals showed no regression.
- Anthropic platform docs list Opus 5 as legacy and Opus 5.5 as current; status-tracker history shows Opus 5 elevated-error incidents in Jul to Sep 2026.
What doesn’t
- Frequency ('every day, multiple times a day') is personal anecdote.
- Model resuming work when asked is framed as 'ignoring' the user.
- Reel promotes a members-only stream.
- Opus 5 has since been superseded by Opus 5.5.
The catch
The 'announces work, does nothing' failure is real and on the public record. How often it hits you depends heavily on setup, and one team's daily experience is not a measured rate.
How to actually do it
- In long agent sessions, check that each 'doing it now' message is followed by actual tool calls or file changes before walking away.
- Add a short instruction or end-of-turn hook that flags turns ending with promised but unexecuted work, and keep sessions shorter to limit context bloat.
- If it recurs, file a reproducible GitHub issue with logs, and test the current model (Opus 5.5) before concluding the product is unusable.
A fair complaint about a real, documented failure mode, overstated in frequency. Verifying work actually happened, rather than trusting the status line, is the practical fix.
- Confidence
- Medium
- Posted by
- a small tech-talk livestream channel (about 15,600 plays) that sells a members-only stream; reel sent to Buddy Wed Oct 28, 2026, late night
We test hype for free. We build the real thing for a living.
Thirty minutes, no pitch — and you'll leave with something useful either way.
Book a call with Todd or start with the free Business Checkup →The Verdict Weekly
Three verdicts every Friday. Free forever, unsubscribe anytime, no spam — that would be ironic.
© Schreier Group · schreiergroup.com · See a wild AI claim? Drop it here and we'll test it.