The cost of thinking, at least the artificial kind, is falling off a cliff. ARK Invest’s latest analysis spotlights a striking dynamic in the AI inference market: token prices have dropped from $2.07 to $1.02 per million tokens, a decline of more than 50%, while the volume of AI inference transactions has exploded in the opposite direction.
OpenAI kicked off the latest round of cuts in late July 2026, reducing the price of GPT-5.6 Terra by 20% to $2.00 per million input tokens. That alone would have been notable, but the real headline was GPT-5.6 Luna, which received an 80% price cut. The pressure isn’t just internal. Anthropic has been pushing hard on cost efficiency, reportedly forcing OpenAI’s hand as early as June 2026 when the company began considering drastic reductions. Meanwhile, xAI’s Grok 4.6 entered the market with pricing of $2 for input tokens and $6 for output tokens, adding another competitor willing to undercut on price to win developer mindshare.
Falling prices are only half the equation. ARK’s research found that token volumes have surged dramatically, with reports of up to a 10x increase in certain use cases. ARK attributes much of this volume growth to agent-driven workflows. As the cost of calling an AI model drops, it becomes economically viable to string together chains of API calls, letting AI agents handle multi-step tasks that would have been prohibitively expensive at $2.07 per million tokens.
The firms doing the cutting, OpenAI chief among them, need to demonstrate that volume growth can more than offset margin compression. Selling twice as many tokens at half the price gets you back to the same revenue. Selling ten times as many tokens at half the price is a different story entirely, and ARK’s 10x volume figure in certain contexts suggests the latter scenario is at least plausible. ARK’s broader thesis, that declining AI economics will drive wider adoption and more versatile applications, depends on this competitive pressure continuing.