The Token Price Index

The Ornn Token Price Index (OTPI) is a daily, volume-weighted blended price per million tokens for each major AI model lab. It lets you track the effective cost of inference across a lab’s mix of production models in one number, rather than tracking each model’s list price by hand.

What the OTPI value means

indexPerMtok is the blended USD price per million tokens for a lab on a given day. It is a volume-weighted blend across the lab’s paid production models, so higher-volume models and providers influence the value more. For example, an anthropic OTPI of 1.00 means $1.00 per million tokens averaged across paid Anthropic inference for that day.

The OTPI uses paid inference activity and provider-weighted posted pricing, not a simple unweighted list price. Treat indexPerMtok as a market-blended cost of serving a lab’s mix of paid models on a given day.

How it’s computed

For each lab, on each settlement day, the index is a volume-weighted blend of posted per-token prices across that lab’s paid production models — effectively the price of the average paid token served for that lab that day. Free tiers and open-weight self-hosting are excluded, and models served in higher volume move the index more than thinly-used ones.

The exact weighting scheme and provider-sampling methodology are maintained by the OTPI team and refined over time. This page describes the current published behavior; if you need the precise formula or a methodology version for research use, contact the team rather than inferring it from values.

Tracked labs

The index is published for these labs: anthropic · openai · google · deepseek · minimax · xiaomi · qwen · moonshotai · z-ai · mistralai · meta-llama Fetch this list programmatically from GET /api/token-types, or the free-tier subset from GET /api/token-types-free. Open-weight models (for example gpt-oss, gemma) and free tiers are excluded. Models without enough pricing coverage on the settlement day are also excluded, so a lab having a light-coverage day is dropped rather than published with a partial blend.

Update frequency and settlement

OTPI values are settled daily, a short time after the day they cover. The lag is deliberate: it lets the day’s activity finish landing before the day is blended and frozen. Once a day is settled its indexPerMtok is stable. Each row includes:

Field	Meaning
`date`	The settlement date (`YYYY-MM-DD`) the value covers
`lab`	One of the tracked labs above
`indexPerMtok`	Blended USD price per million tokens for that lab
`computedAt`	UTC timestamp the value was computed

indexPerMtok is returned at full floating-point precision (for example 1.234567), not pre-rounded. Round in your own presentation layer to whatever precision you display.

Parameters

The OTPI is served from a single endpoint, GET /api/otpi. Without a key, the free tier covers the four public labs — anthropic, openai, google, deepseek — for the trailing 1 month; the other labs and older history require an API key. See Authentication.

Parameter	Type	Description
`date`	`YYYY-MM-DD`	Return that single settlement day. Defaults to the latest settled day when omitted. Ignored when `startDate`/`endDate` are supplied.
`startDate`	`YYYY-MM-DD`	Inclusive start of a settlement-date range. Use together with `endDate`.
`endDate`	`YYYY-MM-DD`	Inclusive end of a settlement-date range. Use together with `startDate`.
`lab`	enum	Filter to a single lab. One of the tracked labs. An unrecognized value returns `400`.

Notes on combining parameters:

Latest vs. specific vs. range. With no date parameters you get the latest settled day. Pass date for one historical day, or startDate+endDate for a window. Do not combine date with startDate/endDate — the API returns 400 if both are present.
lab is optional everywhere. Omit it to get every lab you have access to (the free-4 without a key, or all tracked labs with one) for the selected day(s); include it to get one lab’s series.

Reading the latest values

Omit date to get the latest settled day across the labs you have access to (the free-4 without a key, or all tracked labs with one):

curl "https://api.ornnai.com/api/otpi"

Add lab to get just one lab’s latest value:

curl "https://api.ornnai.com/api/otpi?lab=anthropic"

Pass a specific date to pull one historical day:

DATE=$(python3 -c 'from datetime import date, timedelta; print((date.today() - timedelta(days=7)).isoformat())')
curl "https://api.ornnai.com/api/otpi?date=${DATE}"

Pulling a historical range

To chart the index over time, pass startDate and endDate (both inclusive, YYYY-MM-DD) instead of date. Combine with lab to limit the response to a single lab. Anonymous callers are clamped to the trailing 1 month; pass an API key to reach further back. Compute dates when you make the request so a no-key example stays inside the moving window:

from datetime import date, timedelta
import requests

end_date = date.today()
start_date = end_date - timedelta(days=7)
resp = requests.get(
    "https://api.ornnai.com/api/otpi",
    params={
        "lab": "deepseek",
        "startDate": start_date.isoformat(),
        "endDate": end_date.isoformat(),
    },
)
resp.raise_for_status()
print(resp.json())

A ranged response echoes the requested window back as top-level startDate/endDate (not date), and data holds one row per settled day for the selected lab, oldest first. A single-day query (date, or no date at all) instead echoes back a top-level date. The top-level startDate/endDate echo the window you requested, while the data rows begin at the first day the lab actually has coverage in that window, so the earliest row can be later than the startDate you asked for.

Loading into pandas

The OTPI range response drops straight into a DataFrame for charting or analysis:

from datetime import date, timedelta

import pandas as pd
import requests

end_date = date.today()
start_date = end_date - timedelta(days=7)
resp = requests.get(
    "https://api.ornnai.com/api/otpi",
    params={
        "lab": "deepseek",
        "startDate": start_date.isoformat(),
        "endDate": end_date.isoformat(),
    },
)
df = pd.DataFrame(resp.json()["data"])
df["date"] = pd.to_datetime(df["date"])
df = df.set_index("date").sort_index()

# blended $/Mtok series for the lab, ready to plot
print(df["indexPerMtok"])

See the pandas guide for the same pattern applied to the GPU index.

Coverage and caveats

New labs backfill from their first settlement. Newly added labs return rows only after their first public settlement. Historical ranges that start before that date will include only the labs already published then, so a range’s per-day count can grow over time.
Light-coverage days are omitted, not zeroed. A lab with insufficient same-day pricing coverage is dropped from that day rather than published with a partial or misleading value.
Settled values are stable; the latest day can shift. Because settlement lands a short time after the covered day, the “latest” day advances daily. Pin to a specific date if you need a fixed value.

See the OTPI API reference for the full schema and code samples in every supported language. For the companion index that prices the request rather than the token, see the Workload Cost Index.

Get Started

Concepts

API Reference

Guides

Reference

The Token Price Index

What the OTPI value means

How it’s computed

Tracked labs

Update frequency and settlement

Parameters

Reading the latest values

Pulling a historical range

Loading into pandas

Coverage and caveats

​What the OTPI value means

​How it’s computed

​Tracked labs

​Update frequency and settlement

​Parameters

​Reading the latest values

​Pulling a historical range

​Loading into pandas

​Coverage and caveats

What the OTPI value means

How it’s computed

Tracked labs

Update frequency and settlement

Parameters

Reading the latest values

Pulling a historical range

Loading into pandas

Coverage and caveats