AI inference cost dropped 43% in two months

Summary

AI inference costs have dropped by 43% in less than two months. The cost per token has fallen from $2.04 to $1.18. This decline is driven by aggressive price cuts from OpenAI and Anthropic, the rise of cheap, Chinese-origin open-weight models like DeepSeek V4 Flash, and architectural efficiency gains such as Mixture-of-Experts.

Why it matters

This price collapse is the most significant signal that AI is transitioning from a luxury research tool to a fundamental utility. It is fundamentally changing the economics of developing and deploying AI products. Simultaneously, it increases competitive pressure on leading labs, which must now compete not just on capability but also on cost efficiency. For small-scale companies, the threshold for integrating AI at scale is lowering.

Highlights

• OpenAI has cut prices for the GPT-5.6 family by up to 80%. Anthropic, meanwhile, launched Opus 5 at roughly half the price of its competitor's top-tier model.
• Chinese-origin open-weight models (e.g., DeepSeek V4 Flash) are flooding the market at prices that are a tiny fraction of US frontier models. Chinese labs are also entering the high-end segment simultaneously.
• AI agents that 'think' for hours and chain multiple tool calls are sharply increasing token consumption. This increase partially offsets per-token savings through a Jevons paradox effect, effectively negating them.

Critical take

The headline 43% figure relies on a single data source (Silicon Data via Jefferies) and measures token prices in isolation. The article itself acknowledges that agentic workloads multiply token usage so significantly that the actual bill for a specific task may barely change. This undermines the narrative that 'AI is getting cheaper' for most practitioners. Additionally, the commoditization framework is somewhat teleological; it assumes the current price war is a permanent state rather than a temporary competitive fluctuation. However, this could reverse as the market consolidates.