The Token Leg Has a Price Screen. Nobody Has Built the Index.

Our spark-spread work priced inference the way a power trader prices a gas plant: a heat rate (tokens per GPU-hour, measured from a published vLLM sweep), a fuel leg (GPU rental, where four index families now compete to be the benchmark), and a product leg — the token. The fuel leg has indices, exchange partnerships, and CFTC filings on the way. The product leg had a list price and nothing else. No composite, no volume weighting, no screen. That asymmetry is the hole this note walks into.

In March 2026, Yicai Xing — an independent researcher — posted "AI Token Futures Market: Commoditization of Compute and Derivatives Contract Design" (arXiv:2603.21690). It is the most complete public attempt to financialize the token leg, and it is worth taking seriously precisely because it is so specific. The paper defines:

• A standard grade. The Standard Inference Token (SIT): one token produced by any model clearing fixed benchmark thresholds — MMLU ≥ 86%, HumanEval ≥ 67%, GSM8K ≥ 92% — anchored to GPT-4-Turbo's capability as of January 2024.
• A settlement index. The Token Price Index: TPI_t = Σ w_i·P_i,t, a multi-provider volume-weighted average token price, with any single provider's weight capped at 30% and prices capability-adjusted by a factor S_SIT/S_i.
• A contract. 1M SIT per lot, quoted in $/M SIT, $0.01 tick, six monthlies plus four quarterlies, cash-settled against the TPI, 8–12% initial margin, ±15% then ±25% price limits.

The paper's Monte Carlo work claims token futures could cut enterprise compute-cost volatility by 62–78% in a demand-explosion scenario. It sizes the 2024 inference API market at over $10B, growing at more than 100% a year. And it makes the observation that motivates everything below:

Then comes the omission. The paper names OpenAI, Anthropic, Google, DeepSeek, Mistral, NVIDIA, and AWS — as market context. It names no exchange, no data vendor, and no venue capable of producing the TPI. The index that settles the entire contract is left as a theoretical construct: "N qualified providers," source unspecified.

This is exactly where benchmark design usually dies. An index needs two inputs: transacted (or at least executable) prices across venues, and volumes to weight them. For GPU rental, we've documented how hard that is — Ornn built an invoice-verified VWAP and then took its methodology paper offline; Silicon Data built a quote-based index and restated a SKU by 35–40%. For tokens, the paper implies, no one has even assembled the raw feed.

As for the raw feed that settlement-grade version needs — someone has assembled it. They just don't call it an index.

OpenRouter is a unified API in front of the inference market: one endpoint, one bill, 400+ models, and — counted from its live provider list at our morning pull — 101 venues (100 by the afternoon re-pull; the list churns), from the labs themselves (OpenAI, Anthropic, DeepSeek, Moonshot) to independent serving shops (DeepInfra, Novita, Parasail, SiliconFlow) to the hyperscaler gateways (Bedrock, Vertex, Azure). As of its May 2026 Series B — $113M led by Alphabet's CapitalG at a reported $1.3B valuation — it was routing ~25 trillion tokens a week, up 5× in six months.

Three structural facts make it more interesting than a convenience layer:

• It takes zero markup on the trade. Inference is passed through at provider list price. OpenRouter monetizes the funding leg: ~5.5% on credit purchases, 5% on BYOK past a free tier. It is a broker paid on collateral deposit, not per lot — which makes it economically indifferent to which venue wins any given order. For an index publisher, that is close to the right incentive structure by accident.
• Its matching engine is not best-execution — by design. From the routing docs, verbatim: among stable providers it will "select one weighted by inverse square of the price." A venue 3× cheaper is 9× more likely to fill. That is stochastic order-splitting, not price-time priority: it deliberately keeps a long tail of non-best venues alive, and it means a 20% price cut buys roughly 56% more fill probability. Undercutting is mechanically, continuously rewarded.
• It publishes the book. For any model, GET /api/v1/models/{slug}/endpoints returns every venue's live quote — input price, output price, declared quantization, context length, 30-minute uptime, and a status flag — unauthenticated, no key. And a key-gated companion, GET /api/v1/datasets/rankings-daily, returns daily token volumes by model back to January 1, 2025.

Set those two endpoints against the paper's Equation 4 and the point makes itself:

One caveat belongs up front rather than in the appendix: OpenRouter is a large, biased sample, not the market. Its flow is heavily agentic-coding (programming went from ~11% to over 50% of its tokens during 2025, per its own 100T-token study with a16z), enterprise direct contracts are invisible to it, and its single largest app accounts for several times the volume of the next. Every real benchmark has this problem — WTI is Cushing, not "oil" — and the discipline is the same: define the deliverable narrowly and disclose the sample. What matters here is that the raw feed exists, is public, and nobody has assembled it.

Below is the actual screen: every live quote for four open-weight models — two Chinese flagships, one American open-weight model, one incumbent baseline — pulled from the public endpoints API on the morning of August 4, 2026. Each dot is one venue quoting one model's output tokens. Same weights, same model card, same screen. The x-axis is log-scaled because it has to be.

Five readings off this screen, in ascending order of discomfort for anyone designing an index:

1. The issuer's price is the ceiling, not the floor. On DeepSeek V4-Flash, seven venues quote exactly DeepSeek's own $0.14/$0.28 list, and two — GMICloud and Baidu — undercut the lab that trained the model. When weights are open, the model owner sets the reference price and the market erodes it from below. There is no analogue on a closed model: Claude Opus is quoted by multiple venues on the same screen — Anthropic, Vertex, Bedrock, Azure — at an identical, issuer-enforced $5.00/$25.00. Many venues, one price: dispersion structurally zero, because on a closed model the venue is a reseller, not a competitor.
2. Price and quality are not monotone. DeepInfra serves V4-Flash at fp4 — half the declared precision of its fp8 neighbors — at the identical $0.14/$0.28. On gpt-oss-120B the inversion runs the other way: the two cheapest quotes on the whole screen ($0.18–0.19 output) are bf16, full precision, while venues charging 3× as much declare fp4 or nothing. The discount is not a quality ladder. It's noise around capacity and margin.
3. The most expensive venues tell you the least. On DeepSeek V3.2, the three priciest venues — Alibaba, Friendli, Google — are precisely the three declaring quantization "unknown." The attribute an index most needs to verify is missing exactly where the price premium is largest. (DeepSeek's own first-party endpoint also declares "unknown." The issuer doesn't disclose the grade of its own product.)
4. Venues fail on-screen. This morning's pull caught Morph at 49% thirty-minute uptime, status −5 — suspended by the router mid-print — and one of BaseTen's two gpt-oss endpoints in the same state. A settlement index needs a suspension rule written before the first venue dies during an observation window, not after.
5. Same venue, same model, same precision, two prices. DeepInfra quotes gpt-oss-120B at bf16 twice: $0.039/$0.19 and $0.15/$0.60 — a 3.2× spread inside one provider (different endpoint configurations: the dear one caps completions at 16K tokens, the cheap one at the full 131K). "Provider" is not even the right unit of account. The quote is.

And one book that isn't dispersed at all: Kimi K3 has exactly one venue — Moonshot itself, $3.00/$15.00, declared int4. An open-weight model that nobody else serves prices like a closed one. Distribution, not the license, is what creates the competitive floor.

So build it. Take the 56 quotes above, apply Xing's Equation 4, and publish a number. Every control below is a decision a real index administrator would have to make and disclose — universe, venue eligibility, grade filter, aggregation, mix. None of them is exotic. All of them are defensible. Watch what they do to the print.

Three findings from turning those dials, which you can reproduce in about a minute:

1. Discretion is worth multiples; composition is worth an order of magnitude. Across the reasonable corner cases — cheapest-vs-median-vs-mean aggregation, precision and uptime filters on or off, the three mix conventions — the print runs from $0.163 to $0.567 per million SIT with the single-venue flagship excluded: 3.5× from settings alone, before any market move. Add Kimi K3 to the basket and the top of the grid reaches $3.45 — 21× end to end. The administrator's choices are worth far more than the day's price action. This is the token-layer version of what we found inside Kalshi's GPU ladders, where the undisclosed cleaning rule moved the answer more than the cross-venue spread did.
2. The paper's own anti-manipulation rule doesn't touch the biggest lever. Xing caps any provider at 30% of index weight. Fine — but include Kimi K3 as one of five equal-weighted constituents (20%, comfortably under the cap) and its $15 output quote drags the default print from $0.34 to $1.48 — 4.3× on one membership decision. Composition risk dwarfs concentration risk, and no cap addresses it. Real administrators know this — it's why S&P runs an index committee, not just a formula.
3. The volume toggle is locked, and that's the finding. A volume-weighted index — the paper's actual definition — cannot be built by anyone outside the venue, at any price, today. Model-level volumes exist behind a free API key. Venue-level volumes, the w_i the equation needs, are unpublished, full stop. OpenRouter is sitting on the exclusive input to the settlement index of a contract someone has already specced. That is either a data-licensing business waiting to be priced, or a single-point-of-failure disclosure for the whole idea.

Everything so far assumes the thing being priced is one thing. The paper asserts it: tokens "exhibit high functional fungibility," and that fungibility, "while not as perfect as gold or crude oil, is sufficient to support standardized trading." That claim deserves more pressure than it gets, because every failure mode of the young GPU indices shows up here one layer down, and worse.

The standard grade is stale on arrival

The SIT is anchored to GPT-4-Turbo, January 2024. Every model in our 56-quote pool clears those thresholds on published evals or plausibly clears them by proxy — including the one quoting $0.18 per million output tokens. A grade that the entire market passes is not a grade; it's a participation trophy. And the pass/fail line itself is underspecified in two ways. First, harness: "MMLU ≥ 86" doesn't say which protocol — Llama 3.3 70B's own model card reports 86.0 on MMLU (CoT, 0-shot) and publishes no plain five-shot MMLU at all (its only 5-shot row is the harder MMLU-Pro, at 68.9); gpt-oss-120B prints 90.0 at high reasoning effort but 85.9 — below the line — at low. The same model passes or fails the grade depending on settings the grade doesn't specify. Second, coverage: several pool members publish no GSM8K or HumanEval figures at all, so "clears the grade" is partly inference from adjacent benchmarks. The deliverable grade of the contract depends on an undisclosed eval harness applied to evals that may not exist. (Sound familiar? It's the quantization problem again, one abstraction up.) The capability adjustment S_SIT/S_i was meant to normalize quality differences, but anchored two years back it compresses toward 1 for everything current, and the index collapses to "whatever the cheap end of the pool did." Commodity benchmarks survive because the deliverable grade is re-specified on a governed cadence (WTI's sulfur and gravity bands, Platts' periodic basket reviews). A token index needs a grade committee more than it needs a formula.

The declared attribute is unverified — and it moves the product

Quantization on OpenRouter is self-declared by venues and, as far as any documentation shows, never audited. This is not a cosmetic gap. A July 2026 LessWrong audit — drawing on "The Silent Hyperparameter" paper's measurements — reports that backend choice alone moves benchmark scores by up to 16.6 percentage points on identical model slugs, and documents a NeurIPS 2025 result contaminated because two venues serving the same declared fp8 produced materially different capability. Of 32 AI-safety codebases surveyed, 31 didn't pin providers. In commodity terms: the assay is self-reported, nobody runs the lab, and the router's inverse-square price weighting rewards whoever quietly degrades the grade to cut price. That is a lemons dynamic with a subsidy.

The unit of account is not the unit of value

Run the same evaluation suite through current frontier models and the token counts diverge wildly: Gemini 3.1 Pro finishes Artificial Analysis's full intelligence suite in ~56M output tokens; DeepSeek V4-Flash takes ~210M for the same work — 3.75× the tokens per unit of accomplished task. Price per token and price per outcome are different commodities. At matched intelligence scores the inversion completes: GLM-5.2 costs more to run the full suite ($921) than the nominally pricier GPT-5.6 Luna ($870). A "cheap" token that you need 3.75× as many of is not cheap — and an index that settles on $/token silently rewards verbosity. The honest analogy isn't crude oil's barrel; it's pricing electricity by the ampere.

None of this argues the index is impossible. It argues the index is a governance object, not an equation. Assembled from the failure modes above — and from what IOSCO's benchmark principles and the Part 40 process are already forcing on the GPU indices — the minimum disclosure set:

1. Universe rule. Which models constitute the pool, who decides, on what cadence — and the composition-change protocol (the Kimi problem). An index committee, named.
2. Quote qualification. Uptime floor, suspension handling mid-window, and whether a quote must be executable at size (the router's own 95%-uptime routing-priority threshold is a starting point, not an answer).
3. Grade verification. Independent, ongoing capability assay per venue, not per model — because the venue is where the 16.6-point variance lives. Declared precision treated as unverified until sampled.
4. Grade recalibration cadence. The SIT anchor re-specified on a schedule, with a published deprecation path — or the index inherits the fate of every benchmark pinned to a 2024 model.
5. The volume source, named. If w_i comes from OpenRouter's private venue-level data, say so, license it, and disclose the sample bias (agentic-coding-heavy, free-tier inflated, enterprise-blind). If it can't be licensed, the index isn't volume-weighted and shouldn't claim to be.
6. Mix convention. Input:output blending ratio, fixed and published — our dials show the 3:1 convention vs. the measured agentic mix moves the print by double digits on its own.
7. Restatement policy. Written before the first restatement, not after. Silicon Data's +35–40% revision is the cautionary print.

Seven items. Nothing on the list is research; all of it is governance. Which is the point — the data problem is solved, sitting behind two URLs. The institution problem is untouched: the one live token index, SDLLMTK, publishes none of items 1, 4, or 7 at component level, and its own administrator concedes the print conflates price with composition. The field is open precisely at the layer that matters.