Index methodology

TPPI is now the Tier Price Index: one price per class of model, across labs, for six tiers: Frontier, Production, Open flagship, Everyday, Fast & Quick and Value, as defined for the Follow the Token workshop on 21 September 2026. TTPI follows each lab across its models; TPPI now follows each class across labs.

Each tier is a published list of product lines, each line covering every generation of one product. The workshop named Claude Fable and GPT Astra (Frontier), Claude Opus and GPT Sol (Production), Kimi (Open flagship), Claude Sonnet and GPT Terra (Everyday), Claude Haiku (Fast & Quick), and DeepSeek Flash and GPT Luna (Value). Google, xAI, Qwen, DeepSeek, Z.ai and MiniMax lines are placed by how each lab positions them; the full list is on this page.

Like TTPI since 1.5.0, each tier’s headline is the realized price, spend divided by tokens, from 2026-09-21; a list-price series runs from index inception for tiers made entirely of closed-weight models. Open-weight tiers have no list series, because an open model’s posted price on OpenRouter is whichever host it routes to that day: over 21 September to 6 October that moved Kimi 13% and DeepSeek Flash 40% a day, against 1 to 1.5% for their realized prices. A tier needs members from at least two labs, so it is never one lab’s price.

The Production Price Index series (a chained price index and an average price paid over one production tier) stopped on 2026-10-06. Their settled values were not changed or deleted; they remain in the database as frozen history and are no longer shown or served. Tier history was computed from the stored daily snapshots. TTPI, TCPI and TIS are unchanged.

TTPI’s headline is now the realized price: what buyers actually paid per million tokens, from the USD spend and token counts OpenRouter reports for each model each day. Until now TTPI priced tokens at posted list rates with a fixed 75% input, 25% output blend and no caching. Real routed traffic is about 97% input tokens, and most of that input is served from cache at a tenth of the input price or less, so the list figure ran two to five times above what was paid. On 2026-10-04 Anthropic’s list figure was $7.60 per million tokens; its realized price was $1.09.

The realized series starts on 2026-09-21, the first day OpenRouter reported spend per model. It is not estimated for earlier days. It covers the same models as before (those with a posted token price), uses the same volume rows, variants and coverage floors, and excludes tokens bought with the buyer’s own provider key (BYOK), which OpenRouter does not bill at the model price.

The list-price series continues unchanged, published beside the headline as the list series from index inception (2026-08-15). Its settled history was relabelled from headline to list in migration 0013; no value, day or version on any row changed. Reading the two together shows how much caching and the input-heavy mix take off the rate card.

TIS and Frontier Premium are computed from the list series, as before, so they are unchanged. TCPI and TPPI are unchanged.

TPPI price index restated. It now chains product lines (families) rather than individual models: a line’s price is the usage-weighted price of its generations, so traffic moving to a cheaper new generation of the same line counts as a price cut. Under 1.3.0 a new generation entered at its launch price as no change, so when GPT-6 Luna launched at $0.20 beside GPT-5.6 Luna at $0.45 the index did not move, and over time it could only register price rises on old models. Restated, the index stands at 109.8 on 2026-10-03 where 1.3.0 showed 157.6.

Two families that combined separate products were split, so the index never reads a move between products as a price change: Gemini Flash and Gemini Flash-Lite, GPT mini and GPT nano.

This is a restatement of published values, which the methodology otherwise never does. TPPI had been public for one day under 1.3.0; every TPPI value since inception (2026-08-15) was removed and recomputed from the stored daily snapshots under 1.4.0, recorded in migration 0011. The average price paid is unchanged in value and was re-settled with it, so both series carry one version. TTPI, TCPI, TIS and Frontier Premium are unchanged.

New index: TPPI, the Tokenando Production Price Index. It tracks the production tier (Sonnet, Haiku, Luna, Flash, Gemma, Llama, Qwen Flash and similar) across labs, where TTPI tracks each lab across all of its tiers. Frontier flagships have converged on the same posted price; the production tier is where labs compete on price.

Membership is a curated list of named tiers with a published rule, reviewed monthly, never a price band: a band would drop any model that raised its price, so the index could only ever look cheap. Open-weight families are included. Labs that sell a single tier are out.

TPPI publishes two series. The first is a chained price index (base 100) over closed-weight members that moves only when those models change price. Open-weight members are left out of it, because their OpenRouter price is the cheapest host of the day and flips daily: combined, that host churn moved the index ±15% a day, and an open-only index moved 20-50% a day. An open-weight index is withheld until open models are priced across all their hosts. The second series is the volume-weighted average price paid over all members, the TTPI formula, which also moves when traffic shifts between cheaper and dearer models. Reading them together separates price cuts from adoption of cheaper models.

Existing indices are unchanged: TTPI, TCPI, TIS and Frontier Premium compute exactly as under 1.2.2. TPPI history was computed from the stored daily snapshots back to index inception.

Correction to the 1.2.1 note below: it said the two construction days had token prices captured "roughly a day late". That was accurate for 2026-08-14 (22h57m late) but overstated for 2026-08-15, whose prices were 7h25m late. Posted token prices change on announcement, roughly monthly, so a seven-hour offset does not move the index materially.

Index inception accordingly moved from 2026-08-17 to 2026-08-15, and that day settled. It is the first day with a complete input set: GPU offers captured within the day itself, and volume weights captured the following morning exactly as designed. The late price capture is disclosed on the snapshot record rather than smoothed over.

2026-08-14 remains excluded permanently: its prices were nearly a full day late and no GPU offers were collected at all, so two of the three indices could never settle for it.

Index inception set to 2026-08-17: settlement now refuses any earlier day outright, so the published record has a defined and enforced start.

Values settled during construction on 2026-08-14 and 2026-08-15 were removed (migration 0008). Their token prices were captured roughly a day late, because the end-of-day collector did not exist when those days closed, so the snapshots held the following day’s prices, and the methodology changed three times across them, leaving values the current configuration cannot reproduce. Publishing two days that fail our own reproducibility test, at the exact point a reader looks hardest, was not worth two days of history.

Every raw snapshot from those days was kept, immutable and untouched, including volume observations that can never be re-fetched from upstream. Nothing irreplaceable was discarded, and the settle-run log of what happened during construction was kept as an audit trail.

TIS expanded from one open model on one GPU to five models across nine model-and-GPU combinations (DeepSeek V4 Pro and R1, Kimi K2.6, MiniMax M3 and gpt-oss 120B on H100 SXM, H200 and B200), each at three serving speeds, all sourced from InferenceX comparison pages with per-pairing citations. Readers choose the combination rather than inheriting ours.

Frontier Premium is now restricted to frontier-class open models. gpt-oss 120B is excluded from it: at roughly 60M tokens per GPU-hour it would set the cheapest self-hosted cost by being small and fast, turning the headline ratio into a statement about model size rather than hosting economics. It remains fully available in TIS.

Pairings that answer no real question are no longer settled: an open-weight lab’s API against somebody else’s open model. Only like-for-like (the lab that serves this exact model) and substitution (closed labs, which cannot be self-hosted at all) are published.

Reproducibility checks are now scoped to the methodology version that produced a value. A value settled under an earlier version is frozen history; it cannot be re-derived from today’s constants and is never overwritten or reported as a mismatch. Without this, the first constant change would have made every historical day fail to re-settle, contradicting the purpose of versioning.

Throughput rows now record which lab publishes the open weights, so the index can distinguish a like-for-like pairing (a lab’s own API vs self-hosting its own model, a clean hosting-margin measurement) from a substitution pairing (a closed lab’s API vs switching to a different, open model, which trades away model quality too). The two answer different questions and are now reported separately instead of in one undifferentiated list.

Documented that the implied self-hosting cost assumes full utilisation of the rented GPU and excludes idle capacity, failed requests and operations overhead. It is a floor, not a forecast.

Labelling and presentation only: no computed value changes, and no settled value was altered.

TIS now settles three serving operating points per (model, GPU), namely volume, balanced and latency, instead of one. Published benchmark curves span roughly 11x between the batch and low-latency ends, which moves the Frontier Premium by more than the difference between any two labs; picking one point silently was indefensible, so all three publish and the reader can switch.

The "balanced" point (interactive serving, ~116 tokens/sec per user) is the documented headline: the API default, the citable figure, and the number in the daily summary.

Throughput table reseeded from InferenceX (SemiAnalysis) for DeepSeek V4 Pro 1.6T on B200 at FP4, replacing the unverified DeepSeek V3 / H200 placeholder. InferenceX reports H200 as unmeasured for this model. Operating points are read from an interpolated published curve, disclosed as such.

Weekly staleness check added: verified throughput citations older than 90 days, or whose source URL stops resolving, raise an alert rather than ageing silently.

Initial methodology: TTPI (volume-weighted blended $/MTok per lab, 0.75/0.25 input/output share), TCPI (weighted median $/GPU-hr per GPU from disclosed posted sources), TIS and Frontier Premium (verified-throughput rows only).

Sources: OpenRouter models API (prices), OpenRouter rankings snapshot (volume weights; disclosed with bias), Vast.ai ask book, RunPod GraphQL, Shadeform catalogue (19 clouds).

Drop rules: labs below 90% volume coverage or under 2 priced models drop for the day; GPUs under 3 distinct providers drop for the day. Missing data is never interpolated or carried forward.