Every AI bill is counted in tokens. A token is a piece of a word, and each lab counts and prices its own. That makes the token a billing unit. It does not make it a unit of work.
Token indices have arrived all the same. At least six kinds are published or designed. This piece asks four questions about them.
The evidence uses the token data Ornn settles each day. The Ornn figures in this piece come from Ornn Data under a full research license. The calculations on that data are Kinetic Alpha’s.
- 1. What token indices exist?
- Most price the average token sold or the median list price. A few hold something fixed: the model, the capability or a grade. We found none that publishes the price of a fixed task as a series.
- 2. What would an index of work be?
- A task-based index: a price index for a fixed task. It runs the same job on each model, reads the tokens each model used, and reports dollars per task. A shift of traffic to a cheaper model does not move it. Only a change in the price of the same work does.
- 3. What is a task-based index worth?
- It keeps the price of work apart from the mix of models and the size of requests. Those can move in opposite directions. Anthropic’s average token fell 38% in two weeks while the same models cost 3% more. Google’s average token did not move in four months while the cost of a request rose 70%.
- 4. What does each series do in a risk process?
- A buyer is exposed to the price of its own work. A lab’s suppliers and lenders are exposed to the lab’s revenue per token. A firm that rents chips and sells tokens is exposed to the gap. Each needs a different number, and the average token fits only the second.
Sort them by what each one holds fixed
Every token index answers one question first: when the number moves, what stayed the same? The answers fall into six groups.
| What is held fixed | Index | How it is built | What moves it |
|---|---|---|---|
| Nothing: the average token sold | Ornn Token Price Index [1] Silicon Data LLM Token Expenditure Index [2] | Posted prices weighted by usage. Ornn publishes one number per lab. Silicon Data publishes one for the whole market ($0.81 per million tokens on November 13) and separate open and proprietary series. | Price cuts, new models and any shift of traffic between models. |
| A list of comparable models | BenchLM Token Price Index [3] Pulse Inference Token Index [4] | The median list price. BenchLM takes flagship models at three input tokens to one output token, with March 2023 set to 100; it read 15.5 in October 2026. Pulse takes one open model across its hosts and publishes the median, the low and the spread, each week. | List prices and which models are on the list. No volume. |
| The same models | Ornn Workload Cost Index, like-for-like [5] | Realized cost per request with the shift between models taken out by calculation. | Price and efficiency of the same models. |
| The same capability | Epoch AI price trends [6] Ornn Model Frontier [7] | The cost to reach a fixed benchmark score. Epoch reports it falling about 47% a quarter since 2023. Ornn’s dataset sets cost per benchmark task beside the score, as a snapshot and not a daily series. | The cheapest model that clears the bar. |
| A grade set by rule | Standard Inference Token, in a 2026 contract design [8] | A token from any model that clears fixed benchmark scores, averaged across providers with a cap on each. A design with no venue. | Prices of qualifying models. |
| The same task | None published as a series that we found | A fixed basket of tasks, priced on each model from the tokens it actually uses. Section 02 sets out the idea. | The price of a fixed unit of work. |
A seventh family is the spread: token revenue set against chip rent. Silicon Data calls its version the crunch spread and reports it as the tokens per second a chip must serve to cover its rent. On October 22 that was 1,605 for an H100 on open-model prices [9]. Our own inference spark spread belongs here too.
The groups are not rivals. They answer different questions, and the trouble starts when one is used for another’s job. A study from MIT makes the point with benchmarks: the price of a given level of performance keeps falling, yet the cost of running a benchmark has stayed flat or risen [10]. The token got cheaper and the job did not.
Fix the task, then read the price
The last row of Table 1 is empty because the raw token is a poor unit. Three facts make the case.
- Labs count tokens differently. Anthropic’s newer models split the same text into about 30% more tokens, so two models at one list price bill different amounts for one prompt [11].
- Some tokens are hidden. A model that reasons before it answers bills that reasoning at the output rate and does not return it to the buyer.
- The mix decides the bill. In one published support-agent example, output is 12% of the tokens and 76% of the cost [11]. A blended price per token means nothing unless the mix is fixed.
An index of work turns the problem around. It fixes the job and reads the price. The job is a standard task: a set of prompts with a fixed answer length and fixed settings. The index runs the task on each model, reads the tokens the model actually used, and multiplies by that model’s prices. The result is dollars per task. We call it a task-based index.
The rest follows the way consumer price indices are built. A basket holds several tasks, because a short lookup and a long agent run use tokens very differently. A model must clear a quality bar to be counted. Weights stay fixed for a period. Each period links to the one before, so a model can launch, change price or retire without breaking the series.
View as table
| Model | Provider | List price, input / output, per million tokens | Cost per 1,000 tasks | Normalized price per million tokens |
|---|---|---|---|---|
| Claude Fable 5.1 | Anthropic | $10 / $50 | $39.00 | $17.73 |
| gpt-6-astra | OpenAI | $10 / $50 | $30.00 | $13.64 |
| Claude Opus 5.5 | Anthropic | $4 / $20 | $15.60 | $7.09 |
| Claude Sonnet 5.5 | Anthropic | $2 / $10 | $7.80 | $3.55 |
| gpt-6.1-sol | OpenAI | $2 / $10 | $6.00 | $2.73 |
| Gemini 4 Argon | $2 / $10 | $6.00 | $2.73 | |
| grok-4.7 | xAI | $2 / $6 | $5.20 | $2.36 |
| Gemini 3.8 Flash | $0.75 / $3.75 | $2.25 | $1.02 | |
| DeepSeek-V4-Pro | DeepSeek | $0.80 / $2.39 | $2.07 | $0.94 |
| Claude Haiku 5.5 | Anthropic | $0.10 / $0.50 | $0.39 | $0.18 |
| gpt-6-luna | OpenAI | $0.10 / $0.50 | $0.30 | $0.14 |
Four models list at $2 per million input tokens. On the same task they cost from $5.20 to $7.80 per 1,000 runs. A buyer who compares price sheets sees a tie. A buyer who compares the task sees a 50% spread.
The same result can be shown as one price per token. Divide each task cost by a common token count and the separate input and output prices collapse into a single normalized figure. On that measure the four models run from $2.36 to $3.55 per million tokens.
The property that matters is what such an index ignores. If traffic moves from a dear model to a cheap one, the index does not move, because no price changed. If requests get longer, it does not move, because the task is fixed. It moves only when the same work costs more or less. The same task can also be priced a second way, as the cost to produce it on rented chips. The gap between the two is a margin, and section 04 returns to it.
It keeps the price of work apart from everything else
A bill is three numbers multiplied together: the price of a fixed job, the tokens the job takes, and the number of jobs. An average-token index blends all three with the mix of models. A task-based index reports only the first.
Three episodes from this autumn show what that is worth. The figures come from the token series Ornn settles each day, which section 05 describes [1][5]. In Table 2 the first two columns are Ornn’s data. The last column is what a task-based index would record, reasoned from how it is built. It is not a measured value.
| Episode | Average token | Same models, like for like | A task-based index |
|---|---|---|---|
| OpenAI, weekend of September 12 and 13: volume up 68% | Down 55% against the week before | Not read for these days | No change from the shift itself. A fixed task ignores where traffic goes. |
| Anthropic, October 1 to 14: a cheaper class takes traffic | Down 38% | Up 3% | Little or none. A new model joins the basket as a new item. Ornn’s like-for-like reading suggests the models already in it did not get cheaper. |
| Google, mid-July to mid-November: requests grow 69% | No change | Down 34% | It would follow the price of the same work, as the like-for-like column does. Bigger requests would not move it. |
1. A flood of cheap traffic moves the average. It does not move the work.
On September 12 and 13 OpenAI settled 4.4 trillion tokens a day, against 2.6 trillion the weekend before, and its average price printed $0.17. The week before it had averaged $0.37. Cheap tokens entered the mix and pulled the average down by half. A shift like that does not make any buyer’s existing work cheaper.
2. A new cheap model moves the average. The same models can cost more.
From October 1 to 14 Anthropic’s average price fell from $1.39 to $0.87, or 38%. Ornn’s like-for-like index, which takes out shifts between models, rose 3% over the same days. Daily volume doubled. The Haiku class shows where the traffic went: the cost of a Haiku request fell 48% between the first week and the second, and the tokens in one more than doubled. Haiku 5.5 lists at $0.10 and $0.50 per million tokens [11]. Nobody’s existing work got 38% cheaper. New, cheaper work arrived.
3. Requests get bigger. The token price does not show it.
Figure 3 compares the week to July 20 with the week to November 13. At OpenAI and at Google the tokens in a request rose about 70%. Google’s average token price did not move, yet the cost of a Google request rose 70%. Anthropic went the other way: requests shrank 17% and the cost of a request fell 63%. Three labs show three different stories, and the price of a token tells none of them alone.
View as table
| Mid-July → mid-November | Anthropic | OpenAI | |
|---|---|---|---|
| Average price of a token, per million | $1.97 → $0.87 (-56%) | $1.36 → $0.37 (-72%) | $0.68 → $0.68 (+0.5%) |
| Cost of a request, cents | 7.93 → 2.93 (-63%) | 1.18 → 0.56 (-53%) | 0.32 → 0.55 (+70%) |
| Like-for-like index | 97.0 → 70.4 (-27%) | 88.8 → 24.0 (-73%) | 90.5 → 59.3 (-34%) |
| Tokens in a request, thousands | 40.3 → 33.6 (-17%) | 8.7 → 14.8 (+71%) | 4.8 → 8.1 (+69%) |
What the index is worth
- It tells a buyer what its own work costs. A buyer who tracked its Anthropic bill with the average token would have seen a 38% fall in October while its own work cost 3% more.
- It keeps price apart from behavior. The mix of models and the size of requests are choices a firm makes. The price of the work is set by the market. A budget owner can manage the first two. Only the third calls for a contract term or a hedge.
- It makes sellers comparable. In Figure 1 four models tie on the price sheet and differ by 50% on the task.
- It can settle a contract. Other people’s traffic cannot move a fixed unit of work. An average can be moved by anyone who sends cheap tokens through the venue.
It has limits. A task-based index does not say how much work is being done or what was actually paid. One built on list prices would miss the discounts large buyers get. So it needs two companions: the size of a request and the volume of requests. Used together they explain a bill. The index gives the price of the job, tokens per request gives its size, and volume gives the count.
What each series does in a risk process
Three parties carry token risk, and each needs a different number.
- A firm that buys AI is exposed to the price of its own work. It needs a task-based price index. Until one is published, the nearest tools are the cost of a request, split into price, size and mix, and the median list price.
- A lab’s suppliers and lenders are exposed to the lab’s revenue per token. The average-token index times volume is the right number for them.
- A firm that rents chips and sells tokens is exposed to the gap between the two. It needs token price set against chip rent, and it can hedge the rent side today.
| Series | What it measures | Use it for | Do not use it for |
|---|---|---|---|
| Task-based price index (none published yet) | The price of a fixed unit of work | Deflating a budget. A reference in a supply contract. Comparing sellers. Settling a buyer’s hedge. | Measuring how much work is done. Standing in for a negotiated price. |
| Task-based cost index and margin (none published yet) | The gap between selling tokens and renting chips | Stress-testing a provider’s debt service. Sizing a GPU futures hedge on the cost side. | A buyer’s bill. |
| Average token, by lab (Ornn Token Price Index) | A lab’s average revenue per token | Tracking a lab’s revenue: price times volume. Spotting a shift in mix early. | Hedging a buyer’s bill. It moves with other people’s traffic. |
| Average token, whole market (Silicon Data) | The market’s average price of a token | A barometer. Paired with chip rent, the speed a chip needs to cover its cost. | Settling any one firm’s exposure. |
| Cost of a request, like for like (Ornn Workload Cost Index) | The cost of a request, split into price, size and mix | Explaining why a bill moved. Setting the growth assumption in a budget. | Pricing a fixed job. A request in July and one in November are different jobs. |
| Token volume (Ornn Token Volume Index) | Demand for tokens | A leading sign for utilization and chip rent. | A market total. It is one channel. |
| Median list price (BenchLM, Pulse) | The posted price of comparable models and hosts | Checking a quote. The spread across hosts shows the cost of not shopping. | Anything that needs volume or a fixed job. |
| Fixed capability (Epoch AI) | The long-run price of a level of capability | The price assumption in a multi-year plan. | Today’s price. |
What works today
- A price clause. A supply contract can refer to a number that neither party controls. It needs a published method and inputs that both sides can check. It needs no exchange.
- A split of the bill. A bill’s change is the change in price, plus the change in size, plus the change in mix, plus what is left. Ornn’s workload series gives the first three for three labs. What is left is the buyer’s own basis: its models against the market, and its negotiated price against the list.
- A margin stress test. A lender to a firm that rents chips and sells tokens can shock rent up and token prices down together, and read what remains to pay the debt. The rent side can be hedged now: ICE Futures U.S. lists futures on Ornn’s GPU index [13].
What is still missing
The token side has no hedge. We know of no listed contract that settles on a token price, and section 03 shows why the average token would make a poor one. A contract on it would pay out on other people’s behavior.
A contract that works has to settle on a fixed unit of work, with inputs anyone can check and a calculator who does not sell tokens. That is the case for a task-based index. Until one exists, the toolkit is the three items above, and the first job of a risk desk is to say which of the three exposures it carries.
What Ornn publishes on tokens
The evidence in section 03 rests on Ornn’s token data. Ornn is best known for its index of GPU rent. Since June 2026 it has also settled four token series each day [1]. A venue supplies the inputs. The method is Ornn’s.
| Series | What it measures | Coverage | Access |
|---|---|---|---|
| Token Price Index | The average price of a paid token at each lab, in dollars per million tokens. Each model counts by the tokens it served. | Eleven labs. Daily from June 4, 2026; hourly from November 10. History rebuilt back to July 2025, with no values from April 4 to June 3, 2026. | Four labs free for three months. The rest by subscription. |
| Token Volume Index | Tokens settled each day, growth from a base day, and each closed lab’s share. | Eleven labs counted. Shares for Anthropic, OpenAI and Google. | Subscription |
| Workload Cost Index | The cost of one request in cents, the tokens in a request, and a like-for-like index that takes out shifts between models. | Anthropic, OpenAI, Google and their model classes. From June 11, 2026. | Subscription |
| Model Frontier | Cost per task (token price times tokens per task) beside benchmark scores. | By model. Updated when benchmarks are. | Key required |
Which parts are inputs, and which are Ornn’s own choices
The inputs come from outside Ornn. They are the posted price of each model, the tokens each model served and the number of requests. A venue reports them. Ornn’s documentation names OpenRouter as the source of the hourly series, which is built from that router’s running totals for the day [1]. For the settled daily series the documentation does not name the venue. It calls the inputs venue-reported and says they are not audited payments [5].
The choices are Ornn’s. They are what turns a venue’s feed into an index:
- which tokens count: paid production models only, with free tiers and open models released by closed labs left out;
- the weighting: by tokens served, so a busy model moves the index more than a quiet one;
- the coverage rule: a lab is dropped for a day when too few of its tokens can be priced, and an hourly window prints only when at least 80% are;
- the grouping of models into classes such as Opus, Sonnet and Haiku;
- the like-for-like calculation, which compares seven-day windows, separates the change in the same model’s cost from the change in mix, and links the result day to day.
Two points need care. First, the launch note describes the index as built from paid, executed inference and not from rate cards [16], while the documentation describes a blend of posted prices weighted by volume [1]. Both hold. The volumes are transacted. The prices are the posted prices of the models that served them. On a router the posted price is the price paid. A discount agreed directly with a lab never appears. Second, the exact weighting formula is not public. The documentation says to ask the index team for it [1].
Anthropic’s average fell from $1.81 in mid-August to $0.87 in the week to November 13. OpenAI’s daily value ranged from $0.16 to $1.20. List prices do not change that often. Most of this movement is mix.
The three closed labs settled 2.6 trillion tokens a day in the first week of August and 5.3 trillion in the week to November 13. Anthropic’s share of the three rose from 26% to 39%. Google’s fell from 31% to 15%.
How Ornn’s series and a task-based index differ
Ornn’s like-for-like index and a task-based index aim at the same thing from opposite sides. Ornn reads what was paid and takes the mix out afterwards. A task-based index fixes the work first and then reads the price. Each has what the other lacks.
| Choice | A task-based index | Ornn’s token series |
|---|---|---|
| Unit | A fixed task, in dollars per task | A token in the price index. A request in the workload index. |
| Price | Public list price, captured on a schedule | Posted price of the models that served traffic, each day and each hour |
| Quantities | Measured by running a fixed set of prompts | Reported by a venue: tokens and requests actually served |
| Weights | Fixed for a period | Each day’s tokens |
| A shift to cheaper models | Does not move the index, by design | Moves the price index. Taken out of the like-for-like index by calculation. |
| Bigger requests | Do not move the index. They need their own series. | Raise the cost of a request. Reported as tokens per request. |
| Quality | A bar every model must clear | Not in the price series. A separate dataset sets price per task beside benchmark scores. |
| What it leaves out | How much work is being done, and what was actually paid | What a fixed job costs, since a request in July and one in November are different jobs |
One limit: the sample is one channel. Price times volume gives the dollars behind the index. On November 13 that was about $3.7 million across the three closed labs, on 6.1 trillion tokens. Ornn says the volumes come from third-party sources that lean toward open models [14]. Work bought straight from a lab under contract is unlikely to appear in data a venue reports. Read the series as a clear view of one channel. They are not the labs’ books.
What this rests on
Ornn data. The Ornn Token Price Index, Token Volume Index and Workload Cost Index figures in this piece come from Ornn Data (data.ornn.com) under a full research license. Ornn is the source of that data. The averages, percentages, shares and implied dollar figures built on it are Kinetic Alpha’s calculations and are not Ornn’s figures.
The task example. Figure 1 is an illustration. The token counts for the task (2,000 in, 200 out) and the 30% uplift for Claude models are assumed. A live index would measure each one.
Limits. Ornn’s series were read on November 14, 2026 and entered at four decimal places for prices and 0.1 billion for volumes. Daily token prices are noisy; where the text compares periods it uses seven-day averages, except for the two-week October comparison, which uses the first and last day. List prices come from one compilation and were not checked against each provider’s page. The third column of Table 2 is reasoning, not measurement.
- Ornn Data documentation, “The Token Price Index,” retrieved November 14, 2026. data.ornn.com
- Silicon Data, “LLM Token Expenditure Index,” product page with daily values to November 13, 2026. silicondata.com
- BenchLM, “Token Price Index” and its methodology, retrieved November 14, 2026. benchlm.ai
- Pulse, “Inference Token Index: Methodology v1.0,” retrieved November 14, 2026. pulse-intelligence-exa.pages.dev
- Ornn Data documentation, “The Workload Cost Index,” retrieved November 14, 2026. data.ornn.com
- Epoch AI price trends, as reported by The Decoder, “AI performance costs are falling faster than those of any previous technology.” Second-hand source. the-decoder.com
- Ornn Data documentation, “Data dictionary” and the Model Frontier dataset reference, retrieved November 14, 2026. data.ornn.com
- Yicai Xing, “AI Token Futures Market: Commoditization of Compute and Derivatives Contract Design,” arXiv:2603.21690, March 2026, as reviewed in Kinetic Alpha, The Token Leg Has a Price Screen, August 4, 2026.
- Silicon Data, “The crunch spread: GPU rents and token prices,” with figures to October 22, 2026. silicondata.com
- Hans Gundlach, Jayson Lynch, Matthias Mertens and Neil Thompson, “The Price of Progress: Algorithmic Efficiency and the Falling Cost of AI Inference,” arXiv:2511.23455, November 2025. arxiv.org
- The Frontier Wire, “LLM API pricing compared,” October 15, 2026. A compilation of provider list prices. thefrontierwire.com
- Ornn Data: Token Price Index, Token Volume Index and Workload Cost Index, retrieved under a full research license on November 14, 2026. data.ornn.com
- ICE Futures U.S., “Ornn Compute Price Index (OCPI) B200 Future,” product page. ice.com
- Ornn Data documentation, “The Token Volume Index,” retrieved November 14, 2026. data.ornn.com
- Ornn Data, “Tokens” data page, retrieved November 14, 2026. data.ornn.com
- Alex Wissner-Gross, “The First Frontier AI Token Price Index,” The Innermost Loop. The author states that he helped form and advises Ornn. theinnermostloop.substack.com