Open-Weight Models Could Trigger a Race to the Bottom in AI Pricing

Summary

The article argues that consumer AI subscriptions currently make advanced systems appear inexpensive, while corporate API usage can produce much larger bills. It cites Claude Opus 4.8 at $5 per million input tokens and $25 per million output tokens, with GPT-5.5 described as having the same input price and a $30 output price. The author says OpenAI and Anthropic have so far competed mainly on intelligence rather than price, while older models generally remain at their original prices until they are retired. The result, in the author’s view, is a growing mismatch between AI spending and the value companies receive, including claims that some firms are facing seven-figure annual bills and that Uber used a year of AI budget in four months. The proposed source of pricing pressure is z.ai’s open-weight GLM-5.2, which the article says performs near leading closed models on many benchmarks while costing $1 per million input tokens and $3 per million output tokens in the cloud. Because its weights can be served by many providers or run in a private data center, the model cannot be retired by one vendor and may become cheaper as competition increases; the author notes that OpenRouter providers grew from about five to about 25 within weeks. The article also argues that local deployment could eventually become much cheaper as hardware improves. Its conclusion is an opinion: once open-weight systems are good enough for common work, customers may switch to the cheapest source of sufficient intelligence, forcing at least one major closed-model provider into a broader price competition that could reduce industry margins.