Skip to content

Instantly share code, notes, and snippets.

@edsu

edsu/hn-claude-opus-47.md

Created April 17, 2026 15:11
Show Gist options

  • Select an option

    Learn more about clone URLs
  • Save edsu/558f2a5a60fb04227c9f87b9826eaa24 to your computer and use it in GitHub Desktop.

Select an option

Learn more about clone URLs
Save edsu/558f2a5a60fb04227c9f87b9826eaa24 to your computer and use it in GitHub Desktop.
Download ZIP
Raw
hn-claude-opus-47.md

Conversation: HN Thread on Claude Opus 4.7

User: hey, can you summarize this hacker news thread? https://news.ycombinator.com/item?id=47793411?foo=barlksjlkjslkj

Assistant: Here's the summary of that Hacker News thread (same thread from your follow-up message too):

Article: Anthropic released Claude Opus 4.7, with "adaptive thinking" as its default (and only) thinking mode.

Key discussion themes:

  • Adaptive thinking problems — Developers report the routing mechanism often skips thinking when it shouldn't, and the behavior can't be disabled in 4.7. This is a regression for users who relied on explicit reasoning.

  • Performance degradation — Multiple users see worse results vs. earlier versions. Workarounds exist (forcing high-effort settings, disabling context optimizations) but feel undocumented and hacky.

  • Lack of transparency — Frustration that Anthropic staff acknowledged issues then went quiet, leaving users unsure whether problems are bugs or intentional cost-saving.

  • Product fragmentation — Chat, Code, and Cowork modes have inconsistent capabilities and context handling, making the overall product feel incoherent.

  • Distillation vs. safety tradeoff — Hiding reasoning traces (to prevent competitor distillation) was criticized for undermining the transparency that many considered a safety benefit.

Overall, the thread is fairly critical — lots of experienced users feeling like Opus 4.7 is a step backward in reliability and developer control.

User: it's a long thread, what else are you seeing? any silver linings?

Assistant: More detail:

Silver linings / positive experiences:

  • Long context — arcanemachiner found 4.7 noticeably better than 4.6 at very long conversations (200K+ tokens), where 4.6 would get "dumb."
  • Some users see no regression — heurist tested 4.7 on a classic reasoning test and it passed fine, suggesting adaptive thinking works well for at least some use cases.
  • Verifiable tasks — lambda and FeepingCreature pushed back on doom-and-gloom, arguing that in math/code contexts errors are catchable and dilute rather than compound, so LLM reliability concerns are overstated.

Practical workarounds shared:

  • /effort xhigh in Claude Code to prevent lazy responses
  • Disabling the 1M context window improves performance
  • showThinkingSummaries: true restores thinking visibility
  • Undocumented CLI flag --thinking-display summarized for VS Code wrappers

Interesting tangents:

  • A Conway's Law observation: the product fragmentation (Chat vs. Code vs. Cowork) likely reflects Anthropic's internal org structure more than any technical limitation.
  • A surprisingly lively thread about RLHF and cultural communication norms — someone argued Claude's "eggshells" communication style mirrors Anglosphere politeness norms, with comparisons to Greek, Scandinavian, and Eastern European directness.

Bottom line: The thread is mostly critical, but the long-context improvement is a genuine bright spot, and there's a pragmatic minority reminding people that verifiable/iterative tasks are still very productive with these models.

User: can you save this interaction as a markdown file?

Assistant: Saved to /Users/edsu/tmp/hn-claude-opus-47.md.

User: instead of summarizing can you write our actual conversation transcript to the markdown file?

Assistant: (this file)

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment