Anthropic's Method to Losing Goodwill in a Few Easy Steps

Original Article ↗ Hacker News Discussion ↗

Lock‑in, Pricing, and Subscriptions

  • Big thread focus: Anthropic restricting Claude subscriptions to their own tools (Claude Code, desktop, Slack) and limiting use with third‑party harnesses.
  • Many see this as classic “enshittification”: create dependence with cheap, powerful access, then tighten terms and raise effective costs.
  • Defenders argue it’s a rational loss‑leader: subsidized tokens only make sense if they drive usage of Anthropic’s own software. If you want portability, pay metered API prices.
  • Critics counter that changing terms, model behavior, and quotas mid‑stream erodes trust, especially for people building tooling and workflows around Claude.

Third‑Party Harnesses and Agent SDK

  • Confusion over what’s allowed: CLI claude -p vs Agent SDK vs API, and a “paused” but not canceled plan to move SDK usage to API pricing.
  • Some users have already migrated away because they expect future restrictions and don’t want to rebuild again.
  • There’s frustration that Anthropic’s harness is perceived as inferior while being effectively forced; people would prefer using Pi, OpenCode, Codex, etc. with Claude subs.

Product Quality, Reliability, and UX

  • Repeated complaints about uptime, instability, quota errors, and a buggy Electron desktop app.
  • Others report heavy use despite flaws, arguing the models are strong enough that people accept significant pain.
  • Some note Anthropic’s rapid flip‑flopping on billing and limits as more damaging than any single policy.

Business Model, Subsidies, and Competition

  • Debate over whether subscriptions are truly sold below cost vs just much cheaper than API. No consensus.
  • Some view bundling and subsidy as anti‑competitive; others compare it to normal pricing tactics and market segmentation.
  • Concern that when VC money runs out, all major labs will tighten pricing and access.

Open Models and Alternatives

  • Multiple users report switching to or augmenting with open‑weights (e.g., GLM, DeepSeek, Qwen) via services like OpenRouter or local setups, often at far lower daily cost.
  • Opinion that open models are catching up for 90–95% of tasks, which could commoditize LLMs and undercut closed‑model premiums.

Safety, Guardrails, and Trust

  • Some criticize Anthropic’s “safetyist” posture, downgrades, and guardrails as duplicitous or overreaching; others dispute claims about deliberate wrong answers as unsupported or outdated.
  • Allegations about hidden classifiers and downgrades have made some enterprises wary, per thread links.

Account Bans, Support, and Governance

  • Reports of aggressive bans, opaque appeal processes, and lack of human support, with warnings about depending professionally on such tools.

Meta: Article Tone and Moderation

  • Some call the article a hit piece, citing missing June uptime data and moralized language; others think it accurately captures a pattern of anti‑consumer moves.
  • Noted that the HN post was heavily flagged and moved down the rankings, which some find suspicious.