Index methodology

Raw-first: every collector stores the untouched upstream payload before any parsing; settlement reads only stored raw data, never live sources.

Deterministic and idempotent: re-running a settled day must reproduce every value exactly, or the run fails loudly. Settled values are immutable at the database level.

Drop, don’t fudge: a lab or GPU with insufficient coverage is omitted for the day with a recorded reason. Nothing is interpolated, carried forward, or estimated.

Everything disclosed: offered prices are labeled offered; volume proxies are named with their bias; every methodology change is versioned and logged.

Prices come from the OpenRouter models API, captured at 23:45 UTC as the day’s close; prices are converted to USD per million tokens, and per-model price overrides shadow base prices where present. Free variants and zero-priced listings are excluded — TTPI is a paid-token index.

Volume weights come from the OpenRouter rankings snapshot captured the following morning (05:10/05:40 UTC), restricted to rows upstream attributes to the settled day, variants standard and thinking only (free and batch are excluded). This weight source reflects routed traffic, not any lab’s total global volume — direct enterprise API traffic is invisible to every public index. That bias is the honest cost of an open methodology, and it is disclosed here rather than hidden.

A lab publishes only when it has at least 2 models carrying both a price and volume, and when the priced models cover at least 90% of the lab’s included-variant volume. Below either floor, the lab drops for the day with a recorded reason.

GPU offers are captured four times daily (05:30, 11:30, 17:30, 23:30 UTC) from three disclosed sources: the Vast.ai public ask book (one query per GPU, price-ascending so any truncation at the venue’s 64-offer cap is stable and bites the expensive tail), RunPod’s public GPU pricing (secure and community cloud as two distinct providers), and the Shadeform catalog, which carries posted rates for 19 underlying clouds — each counted as its own provider, only regions with live availability.

All prices are normalised to USD per single GPU-hour (node prices divided by GPU count; Shadeform’s cent-denominated prices divided by 100 — cross-verified against posted dollar rates). Offers outside a per-GPU plausibility band are discarded and counted. Every current source posts asked/on-demand rates, so v1 is honestly an OFFERED-price index; the cleared-source weight is declared now so adding a transaction-based source later is a source addition, not a formula change.

A GPU publishes only with offers from at least 3 distinct providers. Interruptible/spot prices are excluded from the headline and published as a separate companion series. The per-provider median-of-offers followed by a cross-provider weighted median means neither a flood of listings from one venue nor a single outlier provider can move the index.

Throughput constants come from published, citable serving benchmarks, converted as tokens/sec/GPU × 3600 ÷ 10⁶. Each row carries its citation, its serving configuration, and the per-user interactivity it delivers.

TIS publishes only from rows a human has verified against the citation — an unverified row produces no number, ever. Both indices feeding a spread must have settled the same day; a missing input is a drop, never a substitution.

The implied self-hosting cost is a floor, not a forecast: it assumes the rented GPU is busy serving at the cited throughput for every hour you pay for. Real deployments carry idle capacity, failed requests, model loading, and the salary of whoever keeps it running — none of which this index estimates, because none of them are observable market prices. Read the self-hosted side as the best case that disciplined engineering could reach.

TIS currently covers 5 open-weight models across 9 model-and-GPU combinations, each at 3 serving speeds. Model, GPU and speed are all reader-selectable, because each of them moves the cost per million tokens by more than the gap between two labs — fixing any of them silently would amount to choosing the conclusion.

Coverage is bounded by what is publicly benchmarked, not by what we would like to show: a combination appears only when a citable published measurement exists for it, and a GPU appears only when TCPI also prices it that day. Everything else is dropped and logged.

Two different questions hide inside the phrase "cost of serving a token", and blending them would be misleading, so the index reports them separately.

A LIKE-FOR-LIKE pairing puts a lab’s own API against self-hosting that lab’s own open-weight model. The weights are identical and so is the output quality; the only variable is who operates the hardware. That makes the spread a clean measurement of the hosting margin — what you pay for someone else to run the GPUs.

A SUBSTITUTION pairing puts a closed lab’s API against self-hosting a different, open-weight model. Claude, GPT and Gemini weights are not released at any price, so no like-for-like self-hosted comparison for them can exist — that structural asymmetry is precisely what makes the frontier interesting to measure. The spread here bundles a price gap with a capability gap and must be read as a procurement question ("is open-weight good enough for this workload?"), never as a like-for-like saving. The Frontier Premium is a substitution measure by construction.

A third combination — an open-weight lab’s API against self-hosting some other open model — is not published at all: anyone willing to self-host would run the model they already use, so the number would answer nothing.

The Frontier Premium uses only frontier-class open models as its denominator. A small, fast model can produce tokens far more cheaply than a frontier-scale one, so admitting it would make the headline ratio a measure of model size rather than of hosting economics. Models excluded on that basis remain fully available everywhere else in TIS.

A serving stack can be tuned to push many tokens through each GPU, with every user seeing slower output, or few tokens per GPU with every user seeing fast output. Published results for one model on one GPU span roughly 11× across that curve — a wider gap than the one between two different labs — so the operating point you assume decides the answer more than the market data does.

That makes a single silent choice the least defensible act available to this index. So every operating point settles as its own daily series, the dashboard lets you switch between them, and the "Balanced" point is the published default used for citation, the API's default response, and the daily brief. The sensitivity is the finding; hiding it would be the error.

Operating points are read off InferenceX’s published throughput-vs-latency curve; SemiAnalysis interpolates between measured runs, which we disclose rather than present as discrete measurements.

Verified throughput citations are re-checked weekly. A citation older than 90 days, or whose source URL stops resolving, raises an alert rather than ageing silently into the index — but a constant is only ever updated by a reviewed commit with a changelog entry, never automatically.

All days are UTC. For settled day D: GPU offers are captured four times during D; token prices are captured at 23:45 UTC on D; volume weights are captured at ~05:10 UTC on D+1 (the rankings snapshot upstream attributes to D); settlement runs at 06:00 UTC on D+1. Settled values are frozen the moment they are written.

The published record begins at index inception, 2026-08-15. Settlement refuses to produce a value for any earlier day, so the start of the series is a deliberate, enforced boundary rather than wherever data happened to begin. Values computed before inception, while the methodology was still being written, were removed rather than published; the raw observations behind them were kept.

The headline series are free, keyless, and CORS-open. Responses carry the methodology version and coverage metadata on every value.

How prices for individual models are sourced and verified for the Tokenando pricing index is documented separately on the About page.