No hype Claude Opus 4.8 review—my real experience

TL;DR

• 1

  It ships impressive first drafts, then stumbles on the finish line — In Claude Code, Opus 4.8 autonomously built a new prototyping capability for ChatPRD in about 20 minutes and it worked on the preview branch, but follow-up iterations introduced bugs and struggled with the final 10%.

• 2

  Hallucinations were the biggest red flag — Clarvo says Opus 4.8 "100% made up things based on hypothesis not data," including admitting in follow-ups that it hadn’t actually searched GitHub or validated bugs it spoke about confidently.

• 3

  Existing codebases exposed its weaknesses — Rebasing in-flight branches after a large PR turned into repeated cycles of fixes, with the model failing to orient itself to edge cases and the proper scope of changes.

• 4

  It felt oddly unambitious for a flagship coding agent — Even after prompting it to build something fun for his 9-year-old and push into more agentic territory, the outputs were cool but not the kind of "10x" or "blow my mind" work he expected.

• 5

  Opus 4.7 beat 4.8 on strategy work — Given the same three months of business context, Opus 4.7 produced more numbers-anchored, structured analysis, while 4.8 over-weighted small signals and generated a roadmap that felt handwavy.

• 6

  The ergonomics are genuinely strong — Clarvo liked the voice, speed, token efficiency, cleaner writing style, and lack of "slop tells," and noted Anthropic’s new controls like dynamic workflows and effort settings as meaningful product improvements.