Local open models are good enough
Many report Qwen3.8-Flash-Next and similar open models running locally at roughly Opus 4.5–4.8 level, cutting or cancelling their Claude subscriptions. Skeptics say frontier models still win clearly for serious coding.
Anthropic has no real moat
Commenters argue compute and capital are a costly standing army rather than a moat, and research leads evaporate within months as competitors and open weights catch up.
IPO economics look shaky
Debate over whether subscriptions are subsidised versus fat API margins that exclude training costs, plus worries about lockup unlocks, debt-financed compute and regulatory risk.