Every model here is invented
A complete model catalogue, entirely invented
ctxdex is a working catalogue of AI models that do not exist — with the families, snapshots, lifecycle events, provider offerings and pricing structures that real ones have. It exists so you can learn how model catalogues and their pricing actually work, on data you are free to be wrong about.
94 models · 227 provider offerings · Pricing effective
Why synthetic
Real model catalogues change weekly and real prices carry commercial consequences, so neither is safe to teach on. Every model, provider, offering and rate here is authored. Nothing corresponds to a real product and no figure should be quoted as a market price.
What is not invented is the structure. The entity model, the pricing rules and the semantic distinctions are the ones a real catalogue has to get right — which is the part worth learning.
What the catalogue holds
94
models
across every family, active through retired
every model recorded, at any lifecycle status
227
provider offerings
13 of them name a routing selector or a snapshot rather than a model
every offering recorded, at any availability status
67
models priced
27 have no current offering — a fact, not a gap
models with at least one offering current on the reference date
1,913
recorded rates
in the providers' own units, never converted
rates reachable from an offering current on the reference date
453
pricing versions
5 of them scheduled, with amounts already published
policy versions on those offerings, including superseded ones
223
pricing rules
bands, premiums, discounts, allowances, minimums
rules on those offerings' policy versions
Scenarios · in design
A real implementation, broken into its priced operations
A pricing calculator answers “what do a million tokens cost”. That is rarely the question. A real piece of work — a support agent handling a ticket, a document pipeline, a voice assistant — is many operations across several models and offerings, each metered differently, each with its own conditions.
- 01
Classify the ticket
A small general model reads the message and routes it. Cheap per call, run on every ticket.
input + output tokens
- 02
Retrieve the account history
Embed the query, search the notes, rerank the hits. Three meters, none of them tokens-in-tokens-out.
embedded tokens · rerank units · search queries
- 03
Draft the reply
A frontier model with the whole thread in context. Long threads cross a context band and cost half again as much.
input + output tokens · long-context band
- 04
Check it before sending
A safety model screens the draft. Its offering records a monthly minimum, so low volume does not mean low cost.
moderated tokens · monthly minimum
The point is not the total. It is seeing that one piece of work touches four meters on three offerings, that the cheapest model per token is not the cheapest step, and where a long-context band or an included allowance changes the answer. This is the surface being designed next.
Three ideas the rest of the site depends on
Most confusion about model pricing comes from three distinctions. They are the reason these pages are laid out the way they are.
Idea one
A price belongs to an offering, not to a model
A model is a thing that exists. An offering is one provider's commitment to serve it, with its own rates, terms and dates. Ask what a model costs and the honest answer is a question back: through whom?
A model served by five providers has five offerings. There are five prices, and none of them is the model's price.
See it in All pricingIdea two
The rate is not what you pay
Recorded rules change the headline figure: long-context bands, tier premiums, commitment discounts, included allowances, volume bands, monthly minimums. One of them can move the real cost by a multiple.
Send more than a provider's context threshold and every charge on that request can be multiplied — the rule sits outside the rate, not in it.
See a priced exampleIdea three
Everything has a date, and two of them
Models are announced, released, deprecated, retired. Prices take effect, get superseded, and are scheduled ahead. A catalogue that shows only the present cannot answer why something changed.
A rate carries the date it was captured and the date it takes effect. Those are different facts, and both are shown wherever a rate appears.
See a model's chronologyWhere to go
Models
Every model, filterable by category, provider, openness and lifecycle. Cards carry one pricing cue and a route to the rest.
Pricing
Every priced model grouped by billing basis, each row opening into its offerings. Then one model, then one offering, in full.
Compare
Up to four models side by side, on capability and on price — with rate rows that refuse to align when the billing basis differs.
Reference
The vocabulary these pages assume, the help notes attached to each surface, and how the catalogue was built.
The shape of the catalogue
What kinds of models
13 categories, so the catalogue exercises meters beyond tokens — images, video seconds, audio minutes, document pages, robot-hours.
- General Purpose39
- Specialized15
- Embedding6
- Image Generation6
- Coding5
- Reasoning5
- Audio Generation3
- Music Generation3
- Reranking3
- Safety3
- Video Generation3
- Speech Recognition2
- Moderation1
Who serves them
5 providers carry current offerings. A model is often served by several of them at several different prices — which is the whole reason pricing lives on the offering, not the model.
- Agent Makers60 models
First Party Managed Api
- Nimbrex Cloud52 models
Cloud Managed Model
- Oryventa Models38 models
Hosted Open Model
- Runvora Inference29 models
Serverless Inference
- Astavra Enterprise AI16 models
Enterprise Managed Ai
Authored, not scraped, and not randomised
Authored to be structurally honest
Families have generations, generations have snapshots, models get deprecated and replaced. Offerings start, change status and end. Prices supersede one another and are scheduled ahead. The shapes are the ones a real catalogue must handle.
Deliberately awkward in places
Some models have no current offering. Some tiers have no recorded rate. Some meters are not charged, and others are included in another meter. These cases exist on purpose, because they are where real catalogues mislead people.
Units are never normalised
Rates stay in the provider's own units — per million tokens, per audio minute, per generated image, per document page, per robot-hour. Nothing is converted into a common figure, because that conversion is where false comparisons come from.
Nothing here is ranked
No model is scored, no offering is called cheapest, no list is sorted by price. Across different billing bases there is no common unit to rank in, and within one there are rules that decide the answer.
Start with a model that has 5 prices
Ctxdex Chintana 1 Large is served by 5 providers on 5 offerings. One of them — Oryventa Models — does not meter requests at all: it prices in accelerator instance hour. It is the shortest route to understanding why a model does not have a price.