06-reference

innermost loop frontier ai token price index

2026-06-16·reference·source: Innermost Loop·by Alex Wissner-Gross
ai-pricingtoken-costsfrontier-modelscapital-cyclesinference-economicsornnprice-discoverycommodity-markets

Why this is in the vault

The Ornn Token Price Index (OTPI) is the first benchmark to price frontier-lab inference tokens from real transacted volume rather than posted rate cards. For RDCO — a solo-founder shop that runs Claude API daily and tracks AI cost/capability curves for agent infrastructure decisions — a live, volume-weighted daily index for Anthropic and OpenAI token pricing is a direct input to build-vs-buy modeling, agent cost forecasting, and the broader investing thesis around AI compute capital cycles. This is also a companion to the existing [[2026-04-29-dwarkesh-reiner-pope-gpt5-claude-gemini-training]] inference economics reference — Pope explains why pricing looks the way it does; OTPI prices the output side in real time.

The core argument

The AI economy has long had a price for compute (GPU time) but never a price for the token — the unit the economy actually sells. Wissner-Gross draws the analogy to oil: before 1866, crude moved in barrels of inconsistent size and had no standardized price. The token already cleared the first gate (labs bill by it universally); OTPI is the second gate — price discovery from actual transactions.

Posted rate cards are not transaction prices. Caching, input/output asymmetry, provider routing, and model mix mean realized cost diverges structurally from the rack rate. OTPI weights every model by transacted token volume into a single daily figure in dollars per million tokens for each of the two leading frontier labs (Anthropic and OpenAI). Because the index is volume-weighted from hundreds of billions to trillions of tokens daily, it carries a signal labs guard closely: how traffic actually splits across models, and how fast a new release captures share after launch.

The pairing with the Ornn Compute Price Index (OCPI) — already on the Bloomberg Terminal and ICE futures — means both sides of the AI cost curve now have live benchmarks: compute in (OCPI) and tokens out (OTPI). The gap between what a token costs to produce (falling ~40x/year by OpenAI's estimate) and what buyers pay is the real deflation story — and neither a rate card nor a capability headline can show it. OTPI can.

The historical framing is tight: Fleetwood's 1707 price index averaged without weighting; Laspeyres and Paasche added weights in the 1870s to isolate pure price change. OTPI uses a unit-value index — it lets the live model mix into the number — so it moves with what the market actually pays, not a fixed basket.

OTPI is live at data.ornn.com for Ornn Data subscribers.

Author's disclosure: Wissner-Gross has a financial interest in Ornn (he helped form and advises the company, backed by 021T Capital). The piece carries an explicit investment-disclaimer footer.

Mapping against Ray Data Co

Mapping strength: strong.

Three direct touch points:

  1. Agent infrastructure cost modeling. RDCO runs Claude API for the always-on COO agent. A daily volume-weighted price index for Anthropic tokens is a better cost-basis anchor than the published rate card, especially as caching and model-mix shift with each Claude release. OTPI makes "what is my blended token cost" a measurable number rather than an estimate.

  2. Build-vs-buy and model-routing decisions. The index surfaces which direction token prices are actually moving across the Anthropic and OpenAI model portfolios. If OTPI for one lab trends down faster, that's a signal about where commodity pressure is arriving first — actionable for provider-routing strategy in multi-model agent pipelines.

  3. Investing thesis (Markov capital-cycle tracker). The [[2026-05-27-not-boring-thank-god-for-data-centers]] reference covers the demand side of the datacenter bet. OTPI prices the demand-side output directly: if token prices hold up, the $7T datacenter investment projection has a hard demand signal backing it; if they collapse, the commodity overshoot is visible in real time. This is a direct input to the Markov phase-tracker — [[2026-06-11-stratechery-bajarin-apple-ai-compute]] covers supply lag, OTPI would cover realized demand pricing.

One friction point: OTPI is a paid data product (Ornn Data subscription). The public piece announces the index but doesn't publish the actual numbers. Worth monitoring whether free summary data surfaces via Bloomberg Terminal terminals that Ray has access to through phData, or whether Ornn offers a trial tier.

Related