Inference Token Index
USD / million native tokens
Daily history begins with the first saved observation.
A daily reference for published inference prices.
Explore the market, then look inside each developer.
USD / million native tokens
Daily history begins with the first saved observation.
Only named, eligible models with published contributions receive a weight. Missing models are not treated as having zero market share. Weights sum to 100% within the covered basket.
OpenRouter rankings · Original spending data · Download captured observation
| Model | Reported spend contribution | Index weight | 70/30 price | Price source |
|---|
Changes to a model’s input or output tariff change its blended price. Developer rate cards take priority.
New eligible models and retirements change the basket. An index move is not always a provider price cut.
Developer prices come first. Where an exact, current developer quote is unavailable, a labeled OpenRouter price can enter the index.
No. Both options use posted prices and a fixed 70/30 blend. The geometric option gives eligible models equal weight. The spending-share option weights covered models using OpenRouter’s reported spending contributions; it is not a worldwide market share or an average invoice.
A shared daily observation is collected on the first site request of each UTC date and saved for all visitors. Five-minute polling has been removed. There is no unattended scheduler enabled; days without a visit can have gaps. Source failures are shown, never silently dated as fresh.
Track the wider paid catalog. Inspect exactly which models qualify for the eligible index.
| Model / developer | Input / M | Output / M | 70/30 blend | Price source | Index status |
|---|
Multiple sellers and batch variants do not count as extra models. Distinct dated releases can remain separate. Catalog availability does not itself verify purchase outside OpenRouter.
Your workload inputs affect this calculator only. The official index blend stays fixed at 70/30.
| Model | Per request | Daily | 30 days | Price source |
|---|
Estimates use the base uncached tariff. Context-tier premiums, discounts, taxes, tools, retries, and output limits are not modeled. Check the linked developer terms before purchasing.
A transparent posted-price reference, with every constituent and source visible.
Prices are US dollars per million native tokens. Use standard uncached text pricing at the base context tier. DeepSeek uses its peak tariff; off-peak discounts are excluded. Other conditional tariffs are recorded in each source note.
Each eligible model has equal weight. The same equation is applied separately to each developer. Select the OpenRouter spending-share option to use a weighted arithmetic average instead.
Use OpenRouter’s trailing 30-day task-spending feed. Sum each named model’s contributions across tasks, match exact model identifiers or an unambiguous canonical identifier, and normalize across eligible covered models. Developer indices normalize within that developer. Unknown, other, ambiguous and unmatched identifiers receive no weight. Prices still come from developers first, with OpenRouter only as a price fallback.
The public feed lists task leaders, not every model’s complete spending. Its contributions are partial; coverage is shown with every index. This is a covered-spending proxy, not total AI-market share, token-volume share, or an actual customer bill. Weight changes alone can change the result. No synthetic weights, tokenizer adjustment or quality adjustment are applied.
Source: OpenRouter, as of the displayed capture timestamp; original JSON feed. The feed specifies a trailing window but no period-end timestamp. Its public frontend endpoint can change. On failure, keep the prior dated result and show a warning. One immutable spending observation, including the source payload and selected prices, is stored per UTC date on first request, separately from geometric history. No historical weights are inferred.
The index uses a fixed 70/30 input-output blend and an equal-weight geometric average. It includes eligible paid models across developers and model generations. Adoption and capabilities have not been independently audited for every member.
An eligible member requires positive input and output token prices from a matched developer rate card or the OpenRouter fallback, and no known retirement within 90 days. Known specialist/configuration listings are excluded. All eligible models enter the broad-market reference with equal weights; there is no developer concentration cap. At least six developers are required. OpenRouter pricing does not independently verify purchase from another seller.
Both options use native tokens. Neither adjusts for intelligence, tokenizer differences, successful work, latency or reliability. OpenRouter weights describe covered platform spending only. Composition and source-coverage changes can move the index independently of prices.
Version 1.1 expands eligibility to fallback-priced models and removes the developer cap. Its history is stored separately from version 1.0 to avoid a misleading comparison across methodologies. One immutable observation is saved per UTC date, on the first request that day. The previous dated result remains available on a failed collection. Daily archived files preserve model prices, source timestamps, exclusions and index values. There is no backfilled or simulated index history, and no unattended schedule is currently enabled.
Developer sources are preferred. OpenRouter is the pricing fallback only when a matching developer quote is unavailable. Fallbacks do not establish independent paid availability. A source link, tariff basis and validation timestamp accompany every row.
Understand what each index measures, what it leaves out, and why its value can change.
Calculation methodology · Price sources · Historical evidence
A geometric average of eligible models’ published prices using 70% input and 30% output tokens. Each model receives equal weight.
An arithmetic average weighted by named models’ reported contributions to spending on OpenRouter, normalized within the eligible covered basket.
Each entry below covers both calculation options for that developer. Geometric weights are equal within its eligible models. Spending weights are normalized within its covered models; 100% therefore means the covered developer basket, not 100% of that company’s revenue or total usage.
Comparing developers also compares different model mixes. A developer with more premium, older or specialist offerings can have a different index without being more expensive for the same task. A single covered model makes a spending-weighted sub-index effectively that model’s price.
The video index uses a geometric average of published USD prices per generated second for eligible text-to-video models with native audio. It is separate from the text-token indices.
These are research price references. A derivatives settlement benchmark would additionally need formal governance, revision and fallback policies, independent oversight and controls against manipulation.
First-party rate cards come first. OpenRouter is explicitly labeled whenever it supplies a fallback.
Every selected price links to a developer rate card or an explicitly labeled OpenRouter fallback.
OpenRouter model catalogExplore dated publisher announcements and the observations saved by this site. Historical model events are not a backfilled market index.
Select a point or a table row for its source and pricing basis. Gaps are not interpolated. USD per million native tokens.
| Date | Model / index | Selected price / M | Evidence |
|---|
Historical model-price events and saved index observations are separate records. Daily observations preserve the index value and its constituents on the collection date.
Historical events are independently researched from publisher announcements. Each preserves its exact model version, date basis, units and source. We compute the 70/30 blend, but do not assume a launch price remained unchanged until today or apply today’s basket retrospectively. A complete historical index would also require verified membership, availability, retirements and tariffs on every historical date.
OpenRouter remains a pricing fallback in saved observations. Its current model catalog and creation dates do not establish past prices. Only dated captured quotes are plotted. The model history includes retired models for research; this does not make them current index members.
Published paid API rates per generated second. Text-to-video, native audio included, with matching output resolution.
USD / generated second
Index × output seconds. A cost-equivalent volume, not a promise that each model supports that clip length. Provider duration limits still apply.
One rate per model. Developer sources only in this initial video basket.
| Model | USD / second | Volume cost | Source / availability |
|---|
These have different input workflows, audio settings or extra token charges. Their rates are shown for context and excluded from the headline calculation.
For each resolution, take the equal-weight geometric mean of verified, positive, on-demand USD rates per output second for text-to-video with native audio. Models must have a published paid API listing. No promotional credits, subscription allowances, batch discounts, image/video inputs, editing, upscaling, silent-only outputs or retired APIs enter this basket. Only one seller and one rate per model are counted.
Resolution and audio are controlled, but quality, frame rate, model capability, generation latency and success rates are not adjusted. Regions differ: the Wan quote uses the documented US tariff, xAI uses its published global tariff. This is a small covered basket, not a usage-weighted measure of the entire video market. No video tokens are mixed with text tokens.
The site checks the linked developer rate cards on the first video-index request per UTC day and stores a separate immutable daily video observation. A failed or ambiguous source check retains the previous dated snapshot with a warning. Historical prices are never fabricated. No unattended schedule is enabled. Availability means an official paid listing was verified, not that a paid generation was purchased or that every account or country can access it.