Skip to content
Model Pulse

Token economics

The exchange

A catalog is not a market. This page treats it as one: models grouped into the classes a buyer actually chooses between, gateway quotes measured against the lab's own counter, and an explicit account of how much history exists behind each number. Everything is priced for the chat profile.

Instruments

5 classes · median rate per class

Tracking 7,785 listings gives you no index, because most of them are the same weights sold twice — they collapse to 380 models anyone actually serves. A basket is a class of purchase. Membership comes from verified facts — lab attribution, effective price, context window, licence — never from a model's name, and a class needs 4 constituents before it gets a median.

Open & hostable

64 constituents

$0.165
median / M

Open weights under a commercially usable licence — the buy-or-host decision

dearest
Inkling$1.64

688.2× between the ends of this class

Frontier reasoning

52 constituents

$4.56
median / M

Reasoning-capable, ≥200k context, above $2/M under the current workload

dearest
o1-pro$245.25

115.6× between the ends of this class

Frontier general

4 constituents

$4.46
median / M

≥200k context, above $2/M, no reasoning mode

cheapest
GPT-4.1$3.02
dearest
Sonar Pro$5.89

Workhorse

94 constituents

$0.761
median / M

$0.30–$2/M with at least a 100k context — the default production choice

dearest
GPT-5.4$1.99

6.5× between the ends of this class

Budget

94 constituents

$0.140
median / M

Under $0.30/M — cheap enough that volume, not capability, decides

cheapest
Mistral Nemo$0.022
dearest
Qwen-VL Max$0.295

13.1× between the ends of this class

Basis

Gateway quotes vs the lab's own counter

When a lab sells its own model, every gateway listing is a quote on identical weights, so the gap is pure markup — or a subsidy. Only models with a first-party reference and at least 3 gateway quotes appear; without a lab counter there is no basis to measure, and the cheapest reseller is not a substitute for one. A quote no second venue comes close to is treated as a feed error and kept out of the headline figures.

Below the lab

Proprietary models, undercut

Above the lab

Fattest typical markups

Open weights

Hosts undercutting the lab that trained it

These are not markups being competed away — they are the gap between a lab's managed service and a commodity host running the same public weights. When the weights are public the lab's API is one host among many, so this is ordinary competition, quite unlike a gateway undercutting the only company that has the weights. Read it as a self-host signal, not as evidence the lab is overcharging.

AI equity context

Global listings · one company, many securities

Public-market context sits beside token economics, never inside its rankings. The registry separates companies from their listings, so a company with a local share and an ADR counts once. Exchange sessions are recorded in local time: markets close at different local times and some have a midday break. The page publishes the registry and context only; it does not republish stock quotes.

companies
13
direct and indirect exposure
listings
12
canonical security records
exchanges
12
local trading calendars
private labs
2
no synthetic tickers

Tracked public exposure

  • Microsoftcloud-inference · enterprise-platformMSFT · Nasdaq
  • Alphabetmodel-lab · cloud-inferenceGOOGL · Nasdaq
  • Amazoncloud-inference · enterprise-platformAMZN · Nasdaq
  • NVIDIAcomputeNVDA · Nasdaq
  • Taiwan Semiconductor Manufacturingcompute2330 · Taiwan Stock Exchange
  • Oraclecloud-inference · data-centerORCL · NYSE
  • CoreWeavecloud-inference · data-centerCRWV · Nasdaq
  • Equinixdata-centerEQIX · Nasdaq
  • AccentureservicesACN · NYSE
  • Tata Consultancy ServicesservicesTCS · National Stock Exchange of India
  • A.P. Moller - MaersklogisticsMAERSK-B · Nasdaq Copenhagen

This is market context, not investment advice. Use the listed exchange or company source for current prices. An indirect relationship is never presented as direct ownership.

What we can honestly measure

Coverage

archive
24d
permanent price record
observed
24d
the last 24 days
instruments
5
classes with a median
index
live
publishing
index readiness
24/14

Levels and spreads above are cross-sectional — they need one day of data and are true today. Anything longitudinal (a price index, volatility, a basis trend) needs 14 days of archive, and until then this page says so instead of drawing a line through 24 points.