GLM 5.2 and the coming AI margin collapse
Points and comments are a snapshot, not live.
Open-weight GLM 5.2 matches frontier model quality at 15-20% of inference cost, threatening AI lab margins.
Inference costs, not training, drive AI economics: frontier labs charge $25/MTok at ~90% gross margin. GLM 5.2 costs $4.40/MTok and matches Opus quality on coding tasks, though it lacks vision and web search. Switching is trivial via OpenAI/Anthropic-compatible endpoints. Data privacy concerns with Chinese provider Z.ai can be mitigated by self-hosting open weights. Inference on AMD hardware is 2.75x cheaper per token than Nvidia Blackwell.
Part two will analyze margin collapse effects and industry winners/losers.
What commenters are saying
Commenters split on GLM 5.2's quality vs Opus: some find it inferior on complex tasks but equal or better on well-defined ones and uncensored use cases. Multiple cite Anthropic's increasing refusals on security/anti-cheat modding as a reason to switch. Others argue enterprises will pay premium for support and legal recourse despite cheaper alternatives, citing historical analogues like cloud, office suites, and OS monopolies. Several note GLM 5.2's fast variant at 200-400 tps and vision MCP workaround exist.