OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why?

Original Article ↗ Hacker News Discussion ↗

Rapid, steep price cuts from OpenAI and Anthropic are seen as a sign that large language model inference is becoming a low-margin commodity, squeezed by aggressive Chinese providers and increasingly capable open-weight models. Commenters argue that efficiency gains and better inference optimizations explain some of the drop, but question whether the underlying businesses can ever cover massive training and infrastructure costs, raising the prospect of an AI bubble or eventual consolidation. Many expect incumbents to respond by seeking regulatory advantages, focusing on enterprise contracts and “agentic” assistants, while consumers benefit from cheaper access in the short term.

Commoditization and Price War

  • Many see a classic race-to-the-bottom: LLMs becoming a commodity like bandwidth or airlines, with thin margins and intense price competition.
  • Price cuts are widely attributed to competitive pressure from open-weight models and especially cheaper Chinese APIs (DeepSeek, Kimi, GLM, Qwen).
  • Some argue OpenAI/Anthropic are fighting mainly each other; others say the real threat is “good enough at low cost” from open/local models and third‑party inference providers.

Technical Efficiency and Cost Drivers

  • Several comments point to rapid efficiency gains (KV‑cache tricks, better inference engines, model slimming/distillation) enabling much lower per‑token costs.
  • DeepSeek’s published numbers (high revenue per GPU-hour on a small, optimized model) are frequently cited as proof that inference can be very profitable.
  • Others note these are self‑reported, assume near‑full utilization, and don’t include training/R&D or customer acquisition.

Profitability, Sustainability, and Bubble Concerns

  • Strong disagreement on whether current economics are viable:
    • One camp: inference margins are “bonkers,” allowing 80% price cuts and still positive cashflow.
    • Another: labs are “wildly unprofitable” overall; training and capex dwarf inference profits, and current pricing can’t justify multi‑trillion valuations.
  • Comparisons are made to airlines, telco fiber overbuild, dot‑com, and blockchain: valuable tech but poor or boom‑bust business models.

Regulation, Moats, and Geopolitics

  • Many see the push for AI regulation as attempted regulatory capture: restrict open-weight and foreign (especially Chinese) models to preserve US incumbents’ margins.
  • Debate over feasibility and impact of bans on Chinese/open models: likely to hurt law‑abiding firms more than bad actors, and could hand advantage to other countries.
  • Some expect de facto bans via “certified safe” schemes, compliance requirements, and insurance pressures rather than outright prohibition.

Usage, “Unlimited” Plans, and Demand

  • Skepticism that true “unlimited” token plans can exist, given real energy and compute costs, though some expect mobile‑style “unlimited with soft caps.”
  • Others note potential overbuild: if models become far more efficient and demand saturates, excess GPU capacity may emerge, further depressing prices.