LLM Token Expenditure Index: average cost per million tokens has fallen to 97 cents, part of a sharp months-long decline since hitting a high of $2.07 on May 28
A closely followed measure of artificial intelligence token prices touched fresh lows this week, the latest sign of deflating prices …
Context & Ripple Effects
Token pricing had already become a competitive lever: OpenAI was reported in June to be weighing drastic token-price cuts in anticipation of Anthropic doing the same. The index’s fall reinforces the pricing pressure implied by that report.
For buyers, lower unit prices do not automatically resolve AI-budget strain. Companies had been rationing or tracking AI use after token-heavy workloads blew through budgets, while earlier coverage found reasoning models can consume enough additional tokens to raise total developer costs despite lower token rates.
First-order effects
- Frontier-model providers face a lower revenue yield per million tokens, intensifying pressure to compete on model capability, throughput, or higher-value services rather than list price.
- Enterprise AI teams that track token spending gain immediate budget relief on comparable workloads, though token-heavy reasoning usage can offset part of the saving.
Second-order effects
- OpenAI, Anthropic, and other providers are pushed toward sharper price competition, making the reported expectation of reciprocal cuts more consequential for their inference margins.
- Companies that had imposed token controls can revisit which employee and product use cases clear their budget thresholds, shifting attention from token price alone to cost per completed task.
Third-order effects
- If falling token prices persist alongside rising tokens per request, AI procurement will increasingly be governed by effective workload cost rather than the advertised token rate.
- The pattern points to AI-margin inversion: cheaper raw inference broadens adoption while forcing model vendors to find differentiation and revenue beyond undifferentiated token sales.
The trend: AI inference is becoming cheaper per token but more economically complex per useful task as reasoning workloads consume larger token volumes.