FULL STORY

Claude Opus 5 Tops Benchmarks but Sparks Refund Wave

Despite topping benchmarks, Claude Opus 5 triggered a wave of cancellations over real-world performance, while Claude Code growth slumped and watermark disputes deepened Anthropic's reputation crisis.

2026-08-17 ~ 2026-08-21 · 3 episodes · 15 posts

Episode 1 · Users slam Anthropic's direction, say no great model since Opus 4.5 (2026-08-17, 2 posts)

Long-time users express strong disappointment with Anthropic, saying Claude Code felt magical in late 2024 but the product has gone off track for the past six months, with no truly great model since Opus 4.5.

Episode 2 · Claude Opus 5 Tops Benchmarks but Triggers User Exodus (2026-08-19, 10 posts)

Despite leading benchmark scores, Claude Opus 5 has triggered a broad user backlash: reports describe instruction violations, interrupted tasks, verbose output, frequent errors, excessive apologies and unexplained refusals, driving cancellations and raising doubts about pure compute scaling.

Confirmed

  • @gerardsans relayed user reports that Opus 5 defies instructions, breaks tasks mid-stream and shows degraded reasoning, with extremely verbose output even when explicitly told to be concise (e.g., Zinsser style); users called the experience "stupid and frustrating."
  • A development team on Claude Max cancelled, saying that after compute adjustments the AI was "almost entirely useless" — their worst AI experience ever — with frequent errors, forgotten state and ignored Claude.md conventions.
  • User @Ananth7e compiled six recurring complaints: extreme verbosity, over-apologizing and flip-flopping, arguing with instructions, inventing odd phrasings, over-running tasks, introducing hidden bugs that pass tests, and confidently breaking fixes; many felt it was worse day-to-day than Opus 4.8.
  • @repligate and reposter @GhostPilotMD reported that Claude (both Opus 5 and Fable 5) sometimes refuses tasks without warning or explanation, with screenshots; commenters called it a bad sign for Anthropic.
  • User @HystonKayange cancelled over endless apologies for mistakes; @StewartalsopIII posted a screenshot saying he "understands why everyone is cancelling."
  • Critics added that Anthropic neglects core users and kept the 5-hour usage cap, further fueling cancellations.
  • @annetgriffin noted earlier reports that Claude (Fable 5 in her post) had "lost half its IQ" due to three severe Claude Code bugs in August–September, which lost some users; the issues were reportedly fixed before the 2025 holidays.

Why it matters

  • The gap between benchmarks and real-world experience undermines the "high benchmark = good model" assumption.
  • Multiple posters read this as a sign that pure compute scaling is hitting a wall, relevant to the whole industry's roadmap.
  • The cancellation wave shows real churn risk among core developer users; vendors must balance capability gains with usability.

Episode 3 · Claude Code Growth Slows to Single Digits Amid Backlash (2026-08-19, 3 posts)

Claude Code's growth has stalled at 5.2% quarter-over-quarter despite $15.12B ARR, as developer saturation, defections to Codex, and backlash over watermarking and cancellations fuel concerns about decline.