“Claude Code (in particular version 4.6), which does seem genuinely impressive to me and a number of friends I trust, still makes plenty of mistakes, and still can’t be trusted.”
Our reading, the sentence that was checked:
Claude Code (in particular version 4.6), which seems genuinely impressive to Marcus and a number of friends he trusts, still makes plenty of mistakes and still can't be trusted.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Even more good news for the future of neurosymbolic AI, paragraph 20.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-even-more-good-news-for-the-future-e79af0-a1
We recorded 11 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-28. The source may have been edited after 2026-04-12. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Observation
- Deadline, as the speaker wrote it
- None stated. This site does not add one.
- Deadline state when assessed
- No horizon
- Deadline state on 2026-10-29
- No horizon (the release’s reference date)
Conditions
- in particular version 4.6 (scope, stated by the speaker)
- seems genuinely impressive to him and friends he trusts (qualifier, stated by the speaker)
From the speaker
- Claude Code
- in particular version 4.6
- still makes plenty of mistakes
- still can’t be trusted
Our interpretation
- Read 4.6 as the Claude model generation used through Claude Code, rather than the application's version number.
- The plain reading of cannot be trusted, in the surrounding tool-versus-complete-solution discussion, is that consequential coding output still requires verification.
- Plenty means recurring substantive failures, rather than an isolated mistake. Marcus supplies no numerical error threshold.
- Assess the referenced generation's pre-post operating period, February 2026 through April 12, 2026.
- Marcus's and his friends' favorable impressions are subjective context, not conditions that must occur before the reliability claim can be assessed.
Counts as supported if: Direct evaluations of Claude Code using a 4.6 model show repeated incorrect or incomplete coding results, corroborated where possible by documented operational failures.
Counts as contradicted if: Representative evaluations show dependable completion of consequential coding work without recurring substantive errors or a continuing need for verification.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
Outcome: Supported
The dated evidence matches what the claim committed to, under the criteria frozen before grading.
Scored on a plain reading. The speaker named no measure for this claim, so it was judged on the reading an informed reader would take, stated below under our interpretation. Both reviews had to find that reading fair.
- Evidence as of
- Deadline state when assessed
- No horizon
- Assessment recorded
- Policy version
- 2026-10-28.7
Rationale
Supported under the contextual reading that Claude Code remained a useful tool whose consequential output needed verification. A direct Opus 4.6 evaluation found repeated final task failures, and Anthropic separately confirmed operational problems during the period. The evidence does not support a stronger claim that every use is unreliable or that the tool cannot be useful. Successful benchmark tasks and the April 10 bug fix limit the conclusion.
Limitations
- The controlled evaluation covers one programming language and a particular task suite.
- Some failed tasks were nearly correct.
- The operational bug was fixed before publication of Marcus's post.
- The result depends on interpreting trust as unattended reliability.
Evidence (2 items)
Supports the claim: Zhang et al. / arXiv,
claude-opus-4.6 solves 15/22 (68.2%)
Why it matters here: The study uses Claude Code with Opus 4.6 and finds multiple final submissions that fail its correctness tests. It also reports strong performance on easy tasks and notes that many failures are near-complete solutions, limiting how broadly the failure rate can be generalized.
- Publisher
- Zhang et al. / arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Supports the claim: Anthropic,
forgetfulness, repetition, and odd tool choices
Why it matters here: Anthropic traces these behaviors to a Claude Code bug introduced March 26 that repeatedly discarded reasoning history and affected both Sonnet 4.6 and Opus 4.6. The company says it fixed this specific bug April 10, so it corroborates failures during the period rather than a continuing defect on the post date.
- Publisher
- Anthropic
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Even more good news for the future of neurosymbolic AI
Claude Code (in particular version 4.6), which does seem genuinely impressive to me and a number of friends I trust, still makes plenty of mistakes, and still can’t be trusted.
- Where
- paragraph 20
- Published
- 2026-04-12
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-even-more-good-news-for-the-future-e79af0-a1: Current. Outcome: Supported. Evidence as of 2026-10-28. Public date not recorded.
Recorded events
Assessment publishedBy this site
First assessment published after both reviews and the final-text check passed.
About assessment marcus-even-more-good-news-for-the-future-e79af0-a1.