Silicon Data's LLM Token Spend Index closed Monday at $0.97 per million tokens — the first close below the $1 mark since the benchmark launched late last year, and less than half of this summer's peak. The index shed 8.6% in a single week alone.
The index is listed on Bloomberg Terminal under the ticker SDLLMTK. It tracks more than 200 open-source models and over 120 closed models, drawing on more than 20 separate pricing and trading-volume data sources, and it normalizes for differences in input/output ratios and context-window size across models. Put another way, what it measures is not any single vendor's published list price, but the blended price the market is actually paying, in aggregate, per million tokens.
What's Dragging the Price Down
Three forces are compounding at once. Chinese open-source models such as Moonshot AI's Kimi K3 have pulled down the floor across the entire budget pricing tier. OpenAI cut prices across its GPT-5.6 lineup in late July. And a number of frontier labs have separately rolled out dynamic pricing schemes that sell off idle compute capacity at a discount during off-peak hours.
The first two factors are easy enough to understand. The third is the one that tends to get underestimated. Dynamic pricing means the very same model can sell for different prices depending purely on the time of day it's called, which structurally biases the weighted average downward over time. The index itself is weighted by actual consumption volume, and routing gateways used across the industry spread inference requests across multiple providers automatically — traffic simply flows to whichever provider is cheapest at that moment — and that routing mechanism, on its own, keeps pushing the reading lower.
Revenue Line Squeezed While Compute Commitments Stay Fixed
Charles-Henry Monchau, chief investment officer at Syz Group, put the risk bluntly: "Foundation model labs are the most directly exposed... Token deflation compresses the revenue line while compute commitments stay fixed."
That sentence points straight at where the mismatch actually lives. Data-center leases, GPU purchases and power-supply contracts are all multi-year, fixed-cost commitments, while per-token prices move on a week-to-week basis. Run a rough calculation: if a given provider's inference revenue tracked the index one-for-one, the drop from this summer's peak to today — roughly a halving — would require usage to double just to keep total revenue flat. But doubling usage also means doubling the compute footprint required to serve it, so gross margin doesn't simply snap back to where it used to be. This is only a rough, order-of-magnitude estimate — each company's model mix and discount structure differ widely, so actual outcomes vary from lab to lab.
An Awkward Moment Ahead of the IPO Window
Both OpenAI and Anthropic filed confidentially for public listings with regulators this summer. For companies now preparing investor roadshows, having a per-token pricing metric keep hitting new lows in public view makes for an awkward backdrop — the first question investors are likely to ask will land squarely on pricing power.
There's a slower-moving effect at work here too: cheap prices get remembered. Once individual users and enterprise customers grow accustomed to today's price levels, there's very little room left to raise them back up later. Every additional step open-source models take forward narrows, by that much more, the premium that closed-source vendors are still able to hold onto. The number itself — the index slipping below $1 — carries no special magic in isolation; what it actually marks is that a broader shift in pricing power has already run a meaningful part of its course.
Sources: Silicon Data's index methodology page, Bloomberg Terminal SDLLMTK pricing data, CNBC, CocoLoop, Yahoo Finance. The $0.97 reading, the 8.6% weekly decline and the index's stated coverage all follow Silicon Data's official figures.