LLM prices dropped 80 percent in months and no one is talking about the real reason

OpenAI and Anthropic cut their LLM prices by up to 80 percent in a few months. GPT-4o went from $15 per million input tokens to $2.50. Claude 3.5 Sonnet dropped from $3 to $1. The obvious take is that this is a pricing war. OpenAI trying to undercut Anthropic, Anthropic responding, classic tech bloodbath. Except the timing does not line up. Both companies dropped prices within weeks of each other, sometimes on the same day. That is not competition, that is cost deflation.

The infrastructure caught up

The article points to the real driver: inference got cheaper. Distillation works better than it did a year ago. Quantisation went from research curiosity to production standard. Batch processing for async workloads is now table stakes. All of this compounds. The interesting bit is what this does to the moat. A year ago, if you wanted GPT-4 class reasoning, you paid OpenAI $30 per million tokens or you built your own cluster. Now you pay $2.50 or you run Llama 3.1 405B on rented H100s for roughly the same per-token cost. The gap between hosted and self-hosted is collapsing. This changes the build-versus-buy calculation for anyone running LLMs in production. If you are processing a million requests a day, the API bill just dropped from $450 to $75. That is rent money in Sydney. But it also means the threshold where running your own infrastructure makes sense dropped with it. The decision is no longer about cost, it is about control and latency. I am watching this because the cybersecurity detection models I am building for my thesis depend on embeddings and classification. Right now I use OpenAI embeddings because they are cheap and good. If the price floor keeps dropping, the calculus shifts to whether I want to own the weights and tune the model myself. Six months ago that was not even a question.

---

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