Reference

Glossary, help and about the data

74 terms in 7 groups, 44 help notes grouped by the surface they attach to, and how this catalogue was built.

Terms

Every term reads two ways. The switch changes all of them at once.

Pricing

20 terms

How a rate is written, what it meters, and what changes it.

Allowance

An Allowance is a quantity of usage covered before paid overage charges begin.

Why it matters

A monthly included quota and a permanently free service are different commercial arrangements and produce different costs once the quota is exhausted.

If the first 5,000 grounding queries per month are included, the allowance covers those queries and only eligible overage is charged afterward.

also called included quotaalso called quotaalso called usage allowance

Ancillary Service Pricing

Service Pricing

Ancillary Service Pricing is the separate price for a tool or service used alongside model inference, such as Web Search, File Search, Code Execution, or Grounding.

Why it matters

Separating tool and model charges gives users a correct breakdown and avoids assuming that an available priced service is always used.

A model inference estimate can include a separate Web Search charge only when Web Search is applicable, invoked, and billable for the selected commercial configuration.

also called add-on service pricingsame as service pricingalso called tool pricing

Batch Pricing

Batch Pricing is pricing that applies when a Provider offers distinct economics for Batch delivery.

Why it matters

Batch jobs can have different prices from ordinary request-response usage, but Batch and Economy must not be treated as synonyms.

A Provider can publish a lower Batch rate card for asynchronous job processing while keeping Standard and Economy processing tiers as separate options.

also called batch rate cardalso called batch rates

Billable Quantity

Billable Quantity is the amount of usage that a Rate actually charges after the applicable metering rules have been applied.

Why it matters

Showing billable quantity makes it possible to explain why the number charged by the Provider may differ from the raw value a user entered.

A scenario may contain 600 seconds of audio, while the MeteringProfile converts that into 10 billable minutes.

also called billed quantityalso called chargeable quantity

Billing Meter

Meter

A Billing Meter is the quantity a Provider uses as the basis for charging, such as tokens, images, minutes, pages, requests, or capacity-hours.

Why it matters

Different model and service types are billed in fundamentally different ways, so understanding the meter is essential to interpreting a price.

A video service might bill by output second, while a document service bills by page and a text model bills by tokens.

short for meteralso called pricing meteralso called usage meter

Billing Unit

A Billing Unit is the unit in which a billable quantity is measured, such as TOKEN, IMAGE, MINUTE, PAGE, ROBOT_HOUR, or THROUGHPUT_UNIT_HOUR.

Why it matters

Preserving the original Billing Unit prevents misleading normalization and keeps unlike pricing models comparable without pretending they use the same meter.

An image-generation Rate can use IMAGE as its Billing Unit while a robotics Rate uses ROBOT_HOUR.

also called billable unitalso called pricing unit

Cached Input

Cached Input is input content reused from a Provider cache and, where the Provider publishes separate cache pricing, charged differently from ordinary uncached input.

Why it matters

Caching can reduce cost, but only if cached and uncached portions are partitioned correctly and the Provider's pricing representation is applied once.

With 1,000,000 total input tokens and 300,000 cached tokens, the normal input charge should normally apply to 700,000 uncached tokens and the cache price to 300,000 cached tokens.

also called cache-hit inputalso called cached input tokensalso called cached prompt

Capacity Pricing

Capacity Pricing charges for reserved, provisioned, or dedicated serving capacity rather than only for individual usage events.

Why it matters

Capacity-based economics depend heavily on reserved quantity and utilization, so converting them into a fake universal token Rate would misrepresent the commercial contract.

Ten throughput units billed for 720 hours can be priced as throughput-unit-hours; CTXDEX may separately estimate an effective cost per million tokens for a chosen utilization scenario.

same as capacity-based pricingalso called reserved capacity pricing

Commitment Pricing

Commitment Pricing is pricing that applies only when a specific commercial commitment option is selected or satisfied.

Why it matters

Two commitment options of the same general type can have different durations, covered capacity, discounts, or obligations.

A 12-month capacity commitment can have a different price from a one-month commitment even though both are TERM_COMMITMENT options.

also called commitment ratesalso called committed-use pricing

Historical Pricing

Historical Pricing is pricing that was effective in a past period and is retained so earlier costs can be reconstructed correctly.

Why it matters

Without historical pricing, CTXDEX could not reproduce what a workload would have cost at an earlier date or explain price changes over time.

A Rate effective from January through March remains available for a February estimate even after a new April Rate becomes current.

also called archived pricingalso called historical ratesalso called past pricing

Included Usage

Included Usage is the portion of eligible usage covered by an allowance or quota before additional charges apply.

Why it matters

The same numeric quota can behave very differently if it is per request, per project, per organization, or per month.

A monthly organization-level allowance of 5,000 search queries can cover usage across multiple requests until the monthly pool is exhausted.

also called covered usagealso called included quota usage

Input Token

An Input Token is a token counted from the content sent into a model when the Provider prices that input by tokens.

Why it matters

Input and output tokens often have different Rates, and some of the input may receive separate cached-input treatment.

A request with 1,000,000 input tokens can be priced at the applicable INPUT Rate, with any cached portion partitioned separately when required.

also called input tokensa provider's word prompt token

Long-Context Pricing

Long-Context Pricing is a commercial price change that applies when a workload crosses a Provider-defined context or usage threshold.

Why it matters

A model can technically support a large context window without charging a different price at that size, or a Provider can introduce a price threshold below the technical maximum.

A Model may support 1 million tokens of context while the Provider changes pricing once input exceeds 128,000 tokens.

also called extended-context pricingalso called long-context rates

Metering Profile

A Metering Profile explains how raw workload usage is converted into the exact quantity that a Provider bills.

Why it matters

The same raw workload can be billed differently by different Providers, so CTXDEX keeps usage conversion separate from the monetary Rate.

Two GiB of cache retained for three days can be metered as 6 GiB-days before the storage Rate is applied.

also called metering definitionalso called metering rule

Output Token

An Output Token is a generated token counted as output when the Provider prices model output by tokens.

Why it matters

Providers commonly publish a different price for generated output than for input, so the two quantities must remain separate.

100,000 output tokens priced at 7.20 USD per million output tokens produce a 0.72 USD output charge.

a provider's word completion tokenalso called generated tokenalso called output tokens

Prompt Caching

Caching

Prompt Caching is a Provider mechanism for reusing previously processed input so repeated content may not need to be processed in the same way every time.

Why it matters

Caching can change both workload behavior and cost, but read, write, and storage charges may all have different rules.

A Provider might discount cached reads, charge a premium when content is written to cache, and separately charge for retained cache storage.

also called cache reusealso called prompt cachesame as prompt caching

Rate

A Rate is the price applied to a defined billable quantity, such as 2 USD per 1 million input tokens.

Why it matters

Keeping Rates atomic makes the calculation traceable and lets input, output, services, capacity, and other charge components remain separately explainable.

A Rate of 2 USD per 1,000,000 input tokens applied to 100,000 billable tokens produces a 0.20 USD charge.

also called price ratealso called unit rate

Rate Schedule

A Rate Schedule is one published pricing branch that applies under a particular set of commercial conditions, such as Base, Batch, or Priority pricing.

Why it matters

Rate Schedules let CTXDEX represent alternate published rate cards without treating each alternative as an extra charge.

If Batch delivery has its own published rate card, CTXDEX can select the Batch Rate Schedule instead of the Base Rate Schedule.

same as pricing schedulealso called rate cardalso called rate-card branch

Robot Hour

A Robot Hour is a provider-native billing unit representing one hour of billable robotics compute or robot-serving usage where that unit is used.

Why it matters

Robotics workloads have different economic units from text generation, and preserving those units makes the price meaningful.

7,200 seconds of eligible robotics compute can be metered as 2 ROBOT_HOUR before the robotics Rate is applied.

same as robot-houralso called robotics hour

Throughput Unit Hour

A Throughput Unit Hour is a provider-native billing unit representing one provisioned throughput unit reserved for one hour.

Why it matters

Provisioned capacity is purchased as reserved capacity over time, so its canonical economics should not be rewritten as ordinary per-token usage pricing.

Ten provisioned throughput units held for 24 hours produce 240 throughput-unit-hours of billable capacity.

also called throughput unit-hoursame as throughput-unit-hour

Serving and commercial

13 terms

Who serves a model, on what terms, through which channel.

Capacity Mode

Capacity Mode describes the broad way serving capacity is supplied, such as shared on-demand capacity, provisioned throughput, or dedicated compute.

Why it matters

Capacity mode changes how access is provisioned and can fundamentally change both the calculator inputs and the native billing model.

An Offering may support shared on-demand access as well as a separate provisioned-throughput capacity option.

also called capacity typealso called serving capacity mode

Commercial Agreement

Agreement

A Commercial Agreement is a customer-specific set of negotiated commercial terms with a Provider.

Why it matters

Agreement pricing may apply only to specific charge components or covered products, so it cannot safely be treated as a blanket percentage discount over every charge.

An enterprise agreement could override model input/output charges while leaving Web Search and route fees on public pricing.

also called custom agreementalso called enterprise agreementalso called negotiated agreement

Commercial Offering

Offering

A Commercial Offering is a specific way a Provider makes a Model available under defined commercial and access conditions.

Why it matters

One Model can have multiple Offerings with different Providers, channels, serving options, availability periods, and prices.

A Model can have a first-party managed API Offering and a separate third-party serverless Offering, each with different commercial conditions.

also called commercial availabilityalso called model offeringshort for offering

Commitment

A Commitment is a commercial option in which a customer agrees to a defined term, spend, or other obligation in exchange for specified commercial terms.

Why it matters

Two commitments of the same broad type can have different terms, prices, covered capacity, or effective dates, so CTXDEX resolves the exact option rather than a simple yes/no flag.

A 12-month term commitment on a provisioned capacity option may qualify for a different Rate than the same capacity used without a term commitment.

also called commercial commitmentalso called committed use

Dedicated Compute

Dedicated Capacity

Dedicated Compute is serving capacity reserved for a specific customer or deployment instead of being drawn from a shared pool.

Why it matters

Dedicated capacity can provide isolation or reserved resources, but it has different commercial and utilization economics from shared on-demand usage.

A provider may sell one dedicated accelerator instance for a billing period rather than charging only for tokens processed.

same as dedicated capacityalso called dedicated serving

Delivery Mode

Delivery Mode describes how a request or job is delivered and how results are returned, such as synchronous, streaming, asynchronous, batch, or realtime.

Why it matters

Different delivery modes can affect which serving paths are available and can also select different commercial or pricing branches.

Streaming returns output incrementally, while Batch represents job-oriented processing completed outside an ordinary request-response interaction.

also called delivery typealso called request delivery mode

Processing Tier

Tier

A Processing Tier is a Provider-defined level of scheduling or service treatment, such as standard, economy, priority, or ultrafast processing.

Why it matters

A processing tier can change availability, latency treatment, or price without changing the underlying Model.

A synchronous request could use the Standard tier or, where supported, a Priority tier with different commercial treatment.

also called processing classalso called service tier

Provider

Inference Provider

A Provider is the organization that makes a Model or related service available for use through a commercial or serving channel.

Why it matters

Provider identity can change availability, serving options, access conditions, performance, and pricing even when the underlying Model is the same.

The same open-weight model could be published by one organization and served by several cloud or inference providers.

same as inference provideralso called serving provider

Provisioned Throughput

Provisioned Throughput is reserved serving capacity purchased in provider-defined throughput units rather than relying only on shared on-demand capacity.

Why it matters

Provisioned capacity can offer more predictable throughput but introduces reserved-capacity economics and utilization considerations.

Ten provisioned throughput units billed for 720 hours remain a capacity-time charge even if CTXDEX later calculates an effective cost per million tokens for comparison.

also called provisioned capacityalso called reserved throughput

Publisher

A Publisher is the organization that releases or publishes a Model as part of its model portfolio or technical lineage.

Why it matters

Separating Publisher from Provider lets CTXDEX show who created or published a model even when another organization hosts or sells access to it.

A model may be published by Agent Makers while a separate inference provider offers managed access to that model.

also called model publisheralso called publishing organization

Service

Ancillary Service

A Service is a separately identifiable capability or tool that can be available alongside model inference, such as Web Search, File Search, or Code Execution.

Why it matters

Keeping Service identity separate from model inference prevents tool charges from being assumed merely because a price exists and allows independent service charge components to be explained correctly.

A Web Search Service may be available for an Offering and billed per query, while another Offering may include it or not expose it at all.

also called add-on servicesame as ancillary service

Serving Surface

Surface

A Serving Surface is a particular interface or serving path through which a Commercial Offering can be used.

Why it matters

A single Offering can expose more than one serving path, and those paths can differ in supported modes, tiers, capacity, performance, or access behavior.

An Offering may expose a primary inference API surface that supports synchronous and streaming delivery.

also called inference surfacealso called serving interface

Spend Commitment

Minimum Spend Commitment

A Spend Commitment is an agreement to meet a minimum amount of eligible spending over a defined period.

Why it matters

Treating a monthly spend obligation as a per-request floor would produce incorrect costs and hide the real commercial obligation.

If eligible monthly charges are 8,200 USD against a 10,000 USD commitment, the period-level true-up is 1,800 USD.

also called committed spendalso called minimum spendalso called spend minimum

Models

16 terms

What a model is in this catalogue, and how one is described.

Capability

A Capability is a general ability the model supports, such as tool calling, structured output, or explicit reasoning.

Why it matters

Capabilities help users compare built-in model abilities without confusing them with specific tasks or provider-side services.

Tool Calling can be a model Capability; Web Search can instead be a separately available provider Service.

also called model capability

Context Window

Context

The Context Window is a model-level limit describing how much context the model can work with at one time, where CTXDEX records that characteristic.

Why it matters

A large technical context window tells you what the model can handle, but it does not tell you when a provider may start charging a different long-context price.

A model can support a large context window while the provider's long-context price branch begins at a different threshold.

same as context lengthalso called context limitalso called context size

I/O Representation

Representation

An I/O Representation describes the concrete form of data a model accepts or produces within a broader modality.

Why it matters

Two models can both support the same broad modality but accept or produce different concrete forms of that data.

Image → Raster image and Action → Robot action are examples of moving from a broad modality to a concrete representation.

also called data representationalso called I/O formatsame as input/output representation

Input Modality

An Input Modality is a broad type of data the model is recorded as accepting.

Why it matters

A model can support a modality only as input, only as output, or in both directions, so the direction must be explicit.

A model can accept Text and Image as inputs while producing only Text output.

also called input typealso called supported input modality

Knowledge Cutoff

A Knowledge Cutoff is the latest period the model's built-in knowledge is stated to cover, when that information is defined for the model.

Why it matters

It helps users understand the time boundary of a model's stated built-in knowledge without implying that every model type has such a cutoff.

CTXDEX may display a cutoff as “Oct 2025” when that is the stored precision; another model may have no knowledge-cutoff fact at all.

also called knowledge cutoff datealso called training knowledge cutoff

Modality

A Modality is a broad kind of data a model can accept or produce, such as text, image, audio, video, vector, or action data.

Why it matters

Modality gives a quick view of what kinds of information a model can work with, while representation details prevent that broad label from becoming technically vague.

A model may support Image as an output modality, while its I/O Representation specifies that the output is a raster image.

also called data modality

Model

A Model is the core AI system CTXDEX describes: what it is, what kinds of data it handles, what it can do, and its intrinsic technical characteristics.

Why it matters

Keeping Model facts separate from serving and pricing facts prevents a provider-specific price, tier, endpoint, or latency measurement from being mistaken for an intrinsic characteristic of the AI model itself.

Ctxdex Frontier 4 Pro is a Model. A managed API offering for it, the serving surface used to access it, and the applicable price are separate linked records.

Model Distribution

Distribution

A Model Distribution is a concrete downloadable, self-hostable, or optimized model artifact recorded for a Model or a specific Model Snapshot.

Why it matters

Distribution tells users what artifact is actually available, which is different from a broad openness label and different from accessing the model through a provider-managed CommercialOffering.

An optimized weight package can be recorded as a Model Distribution for a particular Snapshot while a managed API remains a separate CommercialOffering.

also called downloadable artifactalso called downloadable modelalso called model artifact

Model Family

Family

A Model Family groups related models that belong to the same product or technical lineage while still remaining separate models.

Why it matters

A family helps users navigate related models without losing the fact that each member can have different capabilities, modalities, technical characteristics, availability, licences, and pricing.

Several Frontier models can belong to one Frontier family while each keeps its own Model identity and specifications.

also called model family groupingalso called model series

Model Snapshot

Snapshot

A Model Snapshot is a dated, fixed state of a Model that lets CTXDEX refer to what that model was like at a specific release point.

Why it matters

Snapshots make historical comparison and snapshot-pinned commercial availability possible without overwriting the Model's broader identity.

A snapshot dated 2026-01-20 can preserve the model state for that release even if later snapshots add different characteristics.

also called model snapshot versionshort for snapshot

Open Source

Open Source is an openness classification CTXDEX uses for models categorized as open source, while the exact legal permissions remain defined by the model's licence.

Why it matters

It lets CTXDEX communicate the model's broad openness category without replacing the licence or artifact records that users need for actual use decisions.

An OPEN_SOURCE model should still show its specific current primary licence and any recorded distribution artifacts separately.

same as open-sourceacronym OSS

Open Weight

Open Weight is an openness classification CTXDEX uses for a model whose weights are made available under defined terms rather than being represented only as proprietary managed access.

Why it matters

The label helps distinguish the model's openness posture without implying rights, source-code availability, or artifact availability that the catalogue has not separately recorded.

An OPEN_WEIGHT model can have a separate licence record and may or may not have a Model Distribution row available in CTXDEX.

same as open weightsalso called open-weight model

Openness Classification

Openness

Openness Classification is CTXDEX's broad label for how openly a model is published or accessible, such as proprietary, open weight, source available, or open source.

Why it matters

It provides a quick orientation while keeping legal rights, downloadable artifacts, and managed access in their proper separate records.

A model can be classified OPEN_WEIGHT while its licence and any downloadable distribution are shown separately.

also called model opennessalso called openness level

Output Modality

An Output Modality is a broad type of data the model is recorded as producing.

Why it matters

Knowing the output direction tells users whether a model actually produces text, images, audio, actions, or another data type rather than merely accepting it.

A model may accept Text as input and produce Image as output.

also called output typealso called supported output modality

Source Available

Source Available is an openness classification CTXDEX uses when model-related source materials are available under stated terms, without treating that label as the same thing as open source.

Why it matters

It prevents users from assuming that access to source materials automatically means open-source licensing or that downloadable weights are present.

A SOURCE_AVAILABLE model can still require a specific licence and can have no distribution artifact recorded in CTXDEX.

also called source available modelsame as source-available

Task

A Task is a specific kind of work the model is recorded as being able to perform.

Why it matters

Tasks give a more concrete view of model use than broad capabilities and can carry task-specific technical characteristics.

Code Generation and Robot Action Generation are Tasks; Tool Calling is a Capability.

also called model taskalso called supported task

Performance

9 terms

The measures, and what each one is measured against.

First Audio Response

First Audio Response is the delay until the first generated audio from a realtime or speech interaction becomes available.

Why it matters

For conversational audio, hearing the first response quickly is often a better measure of perceived responsiveness than waiting for an entire turn to finish.

A realtime speech surface may begin returning generated audio 190 ms after the relevant input boundary even though the complete conversational turn takes longer.

also called first-audio latencyalso called time to first audio

Latency

Latency is the amount of time between starting an interaction and reaching a defined response milestone, such as the first token, first audio, or completed turn.

Why it matters

Different latency metrics answer different user-experience questions, so comparing them without understanding the milestone being measured can be misleading.

A 180 ms Time to First Token and a 900 ms full response time are both latency measurements, but they describe different points in the interaction.

also called response latencyalso called response time

Measurement Basis

Measurement Basis tells you where a performance value came from, such as a provider-published figure, a measured benchmark, a synthetic reference profile, or a contractual SLA.

Why it matters

Two identical numbers can carry very different evidentiary weight depending on whether they are synthetic demonstration data, measured benchmarks, provider claims, or contractual commitments.

A 200 ms value marked SYNTHETIC_PROFILE is a CTXDEX reference value, while 200 ms marked CONTRACTUAL_SLA represents a contractual target or commitment under the stated conditions.

also called measurement methodalso called performance provenance

Output Tokens per Second

Output tok/s

Output Tokens per Second is the rate at which a text model generates output tokens after generation has started.

Why it matters

A system can start responding quickly but generate slowly, or start more slowly and then generate very quickly; both metrics are needed to understand the experience.

A model surface with 180 ms TTFT and 120 output tokens per second starts quickly and then produces about 120 generated tokens each second under the stated workload.

also called generation speedshort for output tok/salso called tokens per second

Serving Performance

Performance

Serving Performance describes how quickly or efficiently a model or service responds when it is actually being served, using measures such as latency and throughput.

Why it matters

Two offerings of the same Model can feel very different in practice, so performance must be interpreted in the context in which the Model is served.

The same Model can have lower Time to First Token on a priority serving surface than on a standard shared surface without becoming a different Model.

also called inference performancealso called serving metrics

Throughput

Throughput is the rate at which a serving system processes or produces work, such as generated tokens per second or workload units completed per second.

Why it matters

Throughput helps describe sustained processing speed, but the number can change materially with workload size, concurrency, delivery mode, provider infrastructure, and reserved capacity.

A text generation surface may sustain 120 output tokens per second after generation begins, while a document system may instead be measured in pages or documents processed per second.

also called inference throughputalso called processing throughput

Time to First Token

TTFT

Time to First Token (TTFT) is the delay from the start of a text-generation request until the first generated token becomes available.

Why it matters

TTFT strongly affects how responsive an interactive text experience feels even when the total response later takes much longer to finish.

A request can have a TTFT of 200 ms and then continue generating the remaining response at 100 output tokens per second.

also called first-token latencysame as time-to-first-tokenacronym TTFT

Turn Response Latency

Turn Latency

Turn Response Latency is the end-to-end delay for a defined conversational turn, from the turn boundary or input completion to the corresponding response milestone.

Why it matters

Realtime systems can begin speaking quickly but still take longer to complete the intended response, so turn latency captures a different aspect of conversational responsiveness.

A surface might begin audio at 200 ms but have a 450 ms defined turn-response latency under the same reference workload.

also called conversational latencyshort for turn latency

Workload Profile

Performance Workload Profile

A Workload Profile is a fixed test scenario that defines what kind and size of work a performance number was measured or estimated against.

Why it matters

A latency number for a tiny request cannot fairly be compared with one for a long document or high-resolution image unless the workloads are clearly defined.

A text profile might define a 1,000-token input and 250-token output target, while an image profile defines one 1024×1024 standard-quality image.

also called benchmark workloadalso called reference workloadalso called test workload

Lifecycle and dates

8 terms

Every record has dates, and they do not all mean the same thing.

Captured As Of

As Of

Captured As Of tells you the date when CTXDEX captured, verified, or treated a piece of information as its current known reference.

Why it matters

A fact can still be valid but old, or newly captured while describing an earlier effective period, so freshness and effective validity need separate dates.

A comparable-model set can carry an As of date showing when CTXDEX last captured or verified that comparison, even if one of the compared industry models was released much earlier.

short for as ofalso called as-of datealso called captured date

Effective From

Effective From is the date or time from which a recorded fact, price, status, or commercial condition starts to apply.

Why it matters

A price can be announced or captured before it becomes effective, so using the correct start date is necessary for accurate current, historical, and future interpretation.

A new price captured on August 25 can have an Effective From date of October 1 and must not be used for a September estimate.

also called effective datealso called starts onsame as valid from

Effective Until

Effective To

Effective Until is the date or time after which a recorded fact, price, status, or commercial condition no longer applies.

Why it matters

Correct end boundaries prevent expired prices or statuses from being accidentally carried forward into later estimates or displays.

If an old Rate is valid until April 1 and a new Rate starts on April 1, the old Rate applies before April 1 and the new Rate applies from April 1 onward.

same as effective toalso called expires onsame as valid to

Lifecycle Status

Status

Lifecycle Status describes the recorded state of an item in its lifecycle, such as active, preview, deprecated, historical, or upcoming.

Why it matters

A status label helps users understand whether something is currently intended for use, still being previewed, deprecated, historical, or scheduled for the future.

A Commercial Offering can be marked deprecated while its historical pricing remains valid for reconstructing an earlier estimate.

same as lifecycle stateshort for status

Lineage

Lineage describes how one model, version, snapshot, or related item is connected to earlier or later items in its evolution.

Why it matters

Lineage helps users understand how an item evolved without incorrectly assuming that every newer-looking item is a direct successor or derivative.

A newer Model may explicitly record an earlier Model as a predecessor, while another Model in the same family may have no direct lineage relationship.

also called evolution lineagealso called model lineage

Not Applicable

N/A

Not Applicable means the concept or field does not meaningfully apply to the subject being described.

Why it matters

Separating Not Applicable from Unknown lets users distinguish 'there is no meaningful value here' from 'the value should exist but CTXDEX does not know it yet.'

Time to First Token can be Not Applicable to a pure image-generation surface, while an unknown image-generation latency would mean the metric applies but no reliable value is currently available.

also called does not applyshort for N/A

Replacement

Replacement means one item is recorded as taking the place of another item for a defined purpose or commercial path.

Why it matters

A deprecated item does not always have a direct replacement, and a newer item is not automatically the replacement unless the relationship is recorded.

A deprecated Offering may explicitly point to a successor Offering, while another deprecated Offering may simply end without a designated replacement.

also called designated successoralso called replacement path

Unknown

Unknown means CTXDEX does not currently have enough reliable information to state the value.

Why it matters

Treating missing information as a negative fact can mislead users and can corrupt comparisons, filters, and calculations.

If a Provider has not published a model's parameter count, CTXDEX can show it as Unknown rather than assuming the model has no parameters.

also called not knownalso called unknown value

Comparison

3 terms

What may be read across two records, and what may not.

Comparable Model

A Comparable Model is a model selected for comparison because it shares one or more relevant characteristics, tasks, modalities, or positioning dimensions with another model.

Why it matters

Stating that two models are comparable is useful only when users can also see what they are being compared on and what the comparison does not imply.

Two models can be comparable for multimodal generation while differing substantially in context length, deployment model, pricing, or licence.

also called comparison modelalso called peer model

Comparison Basis

Comparison Basis tells you the specific dimension or dimensions on which two models are considered comparable.

Why it matters

A clear basis prevents a narrow similarity—such as image generation—from being mistaken for whole-model equivalence.

A model can be comparable on IMAGE_GENERATION and IMAGE_EDITING without being comparable on reasoning or agentic workflows.

also called basis of comparisonalso called comparison criteriaalso called comparison dimension

Similar Industry Model

A Similar Industry Model is a current real-world model that CTXDEX lists as broadly comparable to a ctxdex model on one or more defined dimensions.

Why it matters

It gives users a recognizable real-world reference point for understanding a synthetic ctxdex model without claiming the two models are the same.

A synthetic reasoning model can list a current real-world reasoning model as broadly comparable on reasoning and coding dimensions, with an explicit 'As of' date.

also called comparable industry modelalso called industry comparablealso called similar real-world model

Calculator

5 terms

The vocabulary an estimate uses. The calculator itself is not built.

Calculation Assumption

Assumption

A Calculation Assumption is a default, derived choice, or contextual condition that materially affects an estimate and should be visible to the user.

Why it matters

An estimate can change because of defaults the user did not explicitly type, so surfacing material assumptions makes the result understandable and reproducible.

The result can state 'Processing tier: Standard (defaulted)' and identify the pricing-effective date used for the selected estimate.

also called estimate assumptionalso called pricing assumption

Calculation Explanation

Explanation

A Calculation Explanation shows how the calculator turned the selected usage and pricing into the displayed cost.

Why it matters

Users can verify why a total changed and understand components such as cached input, services, capacity, or commitments instead of seeing only an unexplained number.

1,000,000 input tokens → metered as 1,000,000 TOKEN → 1.80 USD per 1,000,000 TOKEN → 1.80 USD charge.

also called cost breakdown explanationalso called cost explanationalso called estimate explanation

Cost Estimate

Estimate

A Cost Estimate is the calculator's projected cost for the selected scenario, commercial configuration, usage quantities, and pricing conditions.

Why it matters

An estimate is useful for planning only when users can tell whether it is complete and what assumptions or pricing components it includes.

A complete estimate might show 2.52 USD total with separate 1.80 USD input and 0.72 USD output components.

also called cost projectionsame as estimated costalso called price estimate

Usage Component

A Usage Component is one separately measured part of a scenario, such as input tokens, output tokens, image count, audio duration, pages, or robot compute time.

Why it matters

Keeping usage components separate lets CTXDEX price multimodal and non-token workloads without forcing unlike quantities into one measure.

A multimodal scenario can contain text input tokens, image input count, audio duration, and text output tokens as separate Usage Components.

also called usage inputalso called usage quantityalso called workload quantity

Workload

A Workload is the kind of task or usage scenario a user wants to estimate, such as text generation, image generation, audio, documents, scientific processing, or robotics.

Why it matters

Workload keeps the calculator approachable without hard-coding pricing logic around broad model categories.

Selecting 'Image' can initialize an image-capable Model and Offering, but the actual billing inputs still come from that Offering's pricing and metering structure.

also called use-case workloadalso called workload categoryalso called workload type

Help and methodology

Each note attaches to a surface. They are grouped that way rather than alphabetically, because the question a note answers is nearly always about the page a reader is on.

Reading this site

17 notes

Applies everywhere, on every surface.

Captured As Of

Tooltip

Shows when CTXDEX last captured or verified a time-sensitive reference fact.

Use **Captured As Of** to judge the freshness of information that can change over time.

It records CTXDEX's knowledge or evidence time: when the information was captured, verified, or treated as the current reference. It does **not** tell you when the underlying fact became effective.

For example, pricing can be captured before a future Rate becomes effective, and an industry comparison can be captured at a later date even though the compared models were released earlier.

Effective From

Tooltip

Shows when a fact, price, status, or commercial condition starts to apply.

**Effective From** is the start of an effective-time interval.

For pricing and other effective-dated records, CTXDEX uses the date to decide when the record becomes applicable. A fact can be captured or announced before this date, so **Effective From** must not be replaced with a captured or publication date.

Effective Until

Tooltip

Shows when an effective-dated fact stops applying.

**Effective Until** is the end boundary of an effective-time interval.

Where CTXDEX uses half-open intervals, a record applies up to—but not including—the end instant. A successor record can therefore begin at exactly the same instant without overlap.

An expired price or status must not be silently carried forward as current.

Released

Tooltip

Shows when a Model or Snapshot was released, not when CTXDEX captured the record.

A **Released** date describes the release of a Model or Snapshot.

It is different from **Captured As Of**, which describes CTXDEX's evidence/freshness time, and from pricing effective dates, which describe when a commercial price applies.

CTXDEX keeps these dates separate so release history, data freshness, and commercial validity are not mixed together.

Verified At

Tooltip

Shows when CTXDEX last verified an externally maintained reference record.

**Verified At** is an editorial or source-verification timestamp used for externally maintained reference information.

It records when CTXDEX checked the reference against its supporting evidence. It is not a substitute for the item's release date, captured-as-of date, or commercial effective dates.

Pricing Effective At

Tooltip

Shows the date and time the Calculator uses to choose applicable pricing.

**Pricing Effective At** is the timestamp against which CTXDEX resolves effective pricing.

The Calculator selects the Pricing Policy Version, Rate Schedule, Rates, and related effective-dated pricing that apply at this time. Historical estimates use historical pricing; future estimates can use scheduled future pricing where it exists.

This is different from when CTXDEX captured the pricing data.

Unknown vs Not Applicable

Help

Unknown means the value should be meaningful but is not known; Not Applicable means the concept does not apply.

CTXDEX keeps **Unknown** and **Not Applicable** separate.

- **Unknown** — the concept is meaningful for the subject, but the value is not currently known, verified, or available. - **Not Applicable** — the concept itself does not meaningfully apply to that subject.

Neither state should be collapsed into a dash or treated as the other. This distinction is important for filters, comparisons, performance metrics, and technical characteristics.

Unknown is not zero

Help

Missing or unpublished pricing must not be interpreted as a free price.

A missing or unknown price is **not** the same as a zero Rate.

CTXDEX treats an explicit `0` Rate as valid pricing data only when the commercial pricing authority actually says the charge is zero. If pricing is unpublished, negotiated, unavailable, or missing, the UI should show that state rather than inventing a free price.

This prevents incomplete pricing from producing misleading cost estimates.

Model vs Snapshot

Help

A Model is the continuing identity; a Snapshot is a specific dated or immutable state.

CTXDEX resolves a Snapshot from the Model plus any Snapshot-specific overrides.

```text Model-level facts + Snapshot overrides = resolved Snapshot view ```

A Snapshot therefore does not need to repeat every capability or representation. Missing Snapshot override data must not automatically be interpreted as lack of support.

Publisher vs Provider

Help

The Publisher releases the Model; the Provider serves or commercially offers it.

CTXDEX keeps these roles separate because they describe different responsibilities.

- **Publisher** — creates or publishes the Model in its model portfolio or lineage. - **Provider** — makes the Model available through an inference or commercial serving channel.

The same organization can hold both roles, but a Model can also be published by one organization and served by several Providers.

Model vs Commercial Offering

Help

Model facts describe the Model; a Commercial Offering describes how a Provider makes it available.

A **Model** owns intrinsic technical identity and characteristics.

A **Commercial Offering** describes a Provider-specific way that Model or Snapshot is made available. Provider, access path, delivery mode, processing tier, capacity options, commitments, service availability, lifecycle, and pricing can vary by Offering.

CTXDEX therefore does not treat commercial availability or price as intrinsic Model properties.

Model vs Distribution

Help

A Model is the model identity; a Distribution is a particular downloadable or distributable artifact.

A **Model Distribution** records a way model artifacts are made available for download, installation, or distribution.

It is separate from the Model's identity and from managed inference access. A Model can have no downloadable artifact, one artifact, or several distributions with different formats or conditions.

Artifact availability therefore must not be inferred solely from openness labels.

Openness, Licence, and Distribution are different

Help

Model openness, legal use terms, and downloadable artifact availability are separate facts.

CTXDEX does not use one “open” label to stand in for several different questions.

- **Openness classification** describes the degree or form in which weights/source are available. - **Licence** records the legal terms governing use and redistribution where known. - **Distribution** records whether and how a downloadable artifact is actually made available.

A model can have open weights without being open source, can be source-available under restrictive terms, or can have a licence record without a downloadable distribution. These facts should be shown independently.

Capability vs Service

Help

A Model capability is intrinsic support; a Service is a separately available commercial/tool capability.

A **Capability** describes what the Model itself supports.

A **Service** is a separately identifiable tool or commercial capability made available alongside model inference, such as Web Search, File Search, or Code Execution.

A Service may have its own availability and pricing. The existence of a Model capability does not imply that a separately billed Service is available, and a priced Service does not imply every Offering supports or invokes it.

Context Window vs Long-Context Pricing

Help

A technical context limit and a commercial long-context pricing threshold are different concepts.

The **Context Window** is a technical Model characteristic describing the amount of context the Model can support.

**Long-Context Pricing** is a commercial condition that can change pricing after a Provider-defined threshold is crossed.

The pricing threshold can be lower than the Model's technical maximum, and a large Context Window does not by itself mean long-context pricing exists.

Rate vs Rate Limit

Help

A pricing Rate is a monetary conversion; a rate limit is an operational usage constraint.

In CTXDEX, **Rate** in the pricing model means an atomic monetary price such as currency per token, image, minute, request, or capacity-time unit.

A **rate limit** is an operational constraint such as requests per minute, tokens per minute, or another provider-enforced usage limit.

They can both affect a user, but they answer different questions and must not share the same semantic meaning.

Current vs Historical

Help

Historical records remain available for reference but must not be presented as current availability or pricing.

CTXDEX preserves historical Models, Snapshots, Offerings, and prices so users can understand earlier states and reconstruct earlier costs.

**Current** means applicable now under the relevant lifecycle/effective-date rules. **Historical** means retained for past interpretation.

A historical Offering or expired Rate must stay visibly historical. CTXDEX does not carry it forward as current merely because it is the most recent record available.

Models

9 notes

The model list and a model's own page.

Why performance belongs to serving

Tooltip

Latency and throughput usually describe a serving configuration, not an intrinsic Model property.

Ordinary serving performance can change with Provider infrastructure, Serving Surface, region, delivery mode, processing tier, capacity, concurrency, and workload.

CTXDEX therefore normally attaches latency and throughput reference values to the applicable Commercial Offering, Serving Surface, or capacity configuration rather than to the bare Model.

A Model-level performance value is appropriate only when the metric is genuinely standardized and deployment-independent.

Time to First Token (TTFT)

Tooltip

TTFT measures the delay until the first generated text token becomes available.

Use TTFT to understand **startup responsiveness** for interactive text generation.

It does not describe how fast the rest of the response is generated. Read it together with Output Tokens per Second and the recorded workload/serving conditions.

TTFT values with different workloads, tiers, regions, or Measurement Bases should not be treated as automatically comparable.

Output Tokens per Second

Tooltip

Shows the generated text-token rate after generation has started.

Output Tokens per Second describes the **generation rate after the first token**.

It complements TTFT: one metric describes the startup delay and the other describes ongoing output speed. It is a serving-performance measure, not the number of tokens billed and not a provisioned-capacity unit.

Always interpret it with the stated workload profile and serving conditions.

Performance Workload Profile

Tooltip

Defines the fixed test workload behind a latency or throughput value.

Performance numbers are only interpretable when the workload is known.

A **Workload Profile** records the reference input size, output target, modality-specific characteristics, and other assumptions used for the performance value. It allows metrics to be interpreted and compared within the intended scope.

This performance Workload Profile is different from the Calculator's **Workload** category, which is mainly a discovery/defaulting aid.

Measurement Basis

Tooltip

Shows whether a performance value is synthetic, benchmark-measured, provider-published, or contractual.

The **Measurement Basis** describes how a performance value was obtained or asserted.

CTXDEX can distinguish synthetic reference profiles, benchmark measurements, Provider-published values, and contractual SLA values. This is separate from the statistic itself: for example, P95 describes how the measurements are summarized, while the Measurement Basis describes the evidence source.

Synthetic Profile values are CTXDEX reference/demo data and are not measured external-provider telemetry.

Commercial Offering

Help

Shows a Provider-specific way the Model or Snapshot is commercially available.

A Model can have more than one **Commercial Offering**.

Offerings can differ by Provider, access path, Serving Surface, delivery modes, processing tiers, capacity options, commitments, Services, lifecycle, and pricing. These are commercial/serving facts and do not change the intrinsic identity of the Model.

Pricing and calculator flows therefore resolve through the selected Offering rather than through a bare Model.

Delivery Mode

Help

Describes how a request or job is executed and how results are delivered.

Allowed Delivery Modes come from the authoritative commercial/serving data for the selected Offering and Serving Surface.

Examples can include synchronous, streaming, asynchronous, batch, and realtime delivery, but the UI should display only values actually supported by the selected configuration.

Delivery Mode remains separate from **Processing Tier**, which describes scheduling or service treatment.

Processing Tier

Help

Describes the Provider-defined scheduling or service treatment used for a supported serving path.

Processing Tier is a separate commercial/serving dimension from Delivery Mode.

A tier can affect latency treatment, commercial conditions, and pricing, but only tiers supported by the selected Offering/Serving Surface should be selectable. Changing Delivery Mode can also constrain which Processing Tiers remain valid.

CTXDEX must not infer unsupported combinations simply because both values exist elsewhere in the catalogue.

No current Offering

Help

No current Offering is different from historical-only availability and from having no Offering recorded at all.

CTXDEX keeps commercial availability states distinct:

- **Current Offering available** — at least one applicable current Offering exists. - **Historical or deprecated only** — Offering history exists, but there is no current Offering. - **No recorded Offering** — CTXDEX has no Offering record for the subject.

These states are not Model lifecycle statuses and should not all be rendered as the same empty value.

Pricing

10 notes

All Pricing, a model's pricing, and one offering in full.

How the billing meter works

Help

The billing meter identifies what quantity the Provider actually charges.

A pricing amount is meaningful only together with the quantity it prices.

CTXDEX keeps **metering** separate from the monetary Rate. Raw scenario usage is first converted by the applicable Metering Profile into a **Billable Quantity**. The Rate then values that quantity using the Provider's Billing Unit and denominator.

This is why two Providers can price similar workloads differently—for example by tokens, images, seconds, pages, requests, or a derived storage/capacity unit.

Why CTXDEX keeps provider-native units

Help

CTXDEX preserves the unit and denominator the Provider actually publishes instead of converting everything to tokens.

Provider-native units are part of the commercial meaning of a price.

CTXDEX therefore keeps Rates in units such as tokens, images, seconds, minutes, pages, robot-hours, or throughput-unit-hours when those are the published billing units. It also preserves denominators such as “per 1 million tokens” instead of rewriting every Rate into a tiny per-unit value.

Derived normalized economics can be calculated for analysis, but they are not substituted for the canonical Provider Rate.

Cached input pricing

Help

Cached input is normally a subset of input that can receive different commercial treatment when the Provider supports prompt caching.

CTXDEX separates ordinary input from **Cached Input** so the same tokens are not charged twice.

When cached usage is a subset of total input:

```text uncached input = total input - cached input ```

The cached portion is then priced exactly once using the Provider's published representation—either an explicit cached-input Rate or a cache-read discount/multiplier, not both for the same economic event.

Cache write and cache storage are separate possible charges and should not be assumed to equal cached-read usage.

Batch pricing

Help

Batch is a Delivery Mode and can have a different rate card from ordinary request processing.

**Batch** describes job-oriented delivery/execution. It is not the same concept as an **Economy Processing Tier**.

Where the selected Commercial Offering supports Batch and the Provider publishes separate Batch economics, CTXDEX can select the applicable Batch pricing branch or pricing treatment.

If Batch is not supported by Set-1 commercial availability, pricing data must not make it appear available.

Priority processing pricing

Help

Priority is a Processing Tier that can receive different commercial treatment from Standard processing.

A **Priority Processing Tier** can change scheduling/service treatment and, where the Provider publishes it, can also change pricing.

Priority is separate from Delivery Mode. A request can have a delivery mode such as synchronous or streaming and independently use a supported processing tier.

CTXDEX exposes only tier choices supported by the selected Commercial Offering and applies the pricing representation recorded for that tier.

Allowances and included usage

Help

An allowance covers eligible usage up to a defined amount before paid overage begins.

An **Allowance** is not a permanent zero Rate.

It normally operates over a defined scope and period—for example, the first number of eligible queries per organization per month. **Included Usage** is the portion covered by that allowance. Once the allowance is exhausted, remaining eligible usage can be priced at the applicable overage Rate.

The scope, reset period, and covered usage must therefore be considered when estimating cost.

Commitment pricing

Help

Commitment pricing depends on the exact commercial commitment option, not only a generic committed/not-committed flag.

CTXDEX represents specific **Offering Commitment Options** because commitments can differ in term, covered capacity, effective period, and commercial treatment.

A one-month and a twelve-month commitment may share the same broad type but have different prices. Capacity and commitment can also coexist.

Commitment Pricing remains separate from a customer-specific **Commercial Agreement**, which can contain negotiated terms beyond public commitment options.

Spend commitments

Help

A spend commitment is a period-level minimum-spend obligation, not a per-request minimum charge.

A **Spend Commitment** is evaluated over its defined commercial period and eligible scope.

For example, if an agreement requires a minimum eligible spend for a month, CTXDEX should compare accumulated eligible charges with that obligation and represent any shortfall as a period-level true-up.

It must not apply the monthly commitment as a floor to each individual request or single workload estimate.

Capacity pricing

Help

Provisioned or dedicated capacity is priced in the Provider's native capacity units rather than being forced into token Rates.

Capacity offerings can fundamentally change the billing model.

Provisioned Throughput or Dedicated Compute can be sold using provider-defined capacity units multiplied by time. CTXDEX keeps that **canonical capacity bill** in its native unit.

The Calculator can separately derive workload economics—such as an effective cost per million tokens—when a utilization scenario is supplied. That analytical result is not the Provider's canonical Rate and should not replace it.

Separate Service charges

Help

Tools and ancillary Services can add charge components separate from Model inference.

Services such as Web Search, File Search, Code Execution, Grounding, or other add-ons can have their own pricing authority and billing meter.

CTXDEX keeps these charges separate from Model inference so the estimate can show which component produced each cost.

A published Service price does not by itself mean the Service is supported by the selected Offering, invoked in the scenario, or separately billed. **Available**, **invoked**, and **billed** remain separate states.

Compare

3 notes

Both tabs of the comparison.

Similar Industry Models

Tooltip

These are dated, broad real-world reference comparables—not claims that the models are equivalent.

CTXDEX uses Similar Industry Models to give users recognizable external reference points for a synthetic ctxdex Model.

The comparison is **directional and dimension-specific**. A listed model can be broadly comparable for reasoning, coding, multimodal capability, image generation, robotics, document intelligence, or another recorded Comparison Basis while differing in many other ways.

V1 does not assign a numerical similarity score, and “similar” must not be read as equivalent or interchangeable.

Comparison Basis

Tooltip

Shows the dimensions used to explain why an industry model is considered comparable.

A **Comparison Basis** scopes the comparison.

The recorded dimensions can describe areas such as reasoning, coding, multimodal capability, image generation, audio, video, scientific use, robotics, document intelligence, or retrieval. More than one basis can apply.

Always interpret the comparison through these dimensions and the accompanying rationale rather than as a whole-model equivalence statement.

Comparison As Of

Tooltip

Shows when CTXDEX captured or verified the current industry-comparison set.

Industry models, product positioning, and capabilities change over time, so CTXDEX comparison sets are visibly dated.

The **As Of** date is the comparison set's captured-as-of/evidence date. It tells you when CTXDEX last treated those external references as the applicable comparison set; it is not the release date of every compared model.

Use the date when interpreting or revisiting a comparison.

Calculator

5 notes

Surface not built

The estimate surface, which is not built yet.

Workload

Help

Workload helps you enter the Calculator with useful defaults; it does not decide the price.

The Calculator's **Workload** choice is primarily a discovery and defaulting aid.

It can choose a sensible starting Model, Commercial Offering, and starter scenario. A Model can appear under more than one Workload entry point.

The actual form fields and cost are then derived from the selected commercial/pricing configuration. Workload is therefore not a pricing rule and is different from a Model Task or a performance Workload Profile.

Usage inputs

Help

The Calculator asks only for usage quantities required by the selected Offering and its pricing/metering structure.

Usage fields are dynamic because not every Model or Offering is billed the same way.

A text scenario may need input and output tokens; an image scenario may use image count and quality; audio can use duration; documents can use pages; robotics or capacity can use provider-native time/capacity units. Multimodal scenarios can contain several Usage Components at once.

Model modality or capability does **not** by itself determine which fields are billable.

Estimate status

Help

The result status tells you whether CTXDEX could produce a complete, unambiguous estimate.

Calculator results distinguish at least:

- **Complete** — all required components were priced successfully. - **Incomplete** — required usage or pricing components are missing. - **Ambiguous** — more than one unresolved pricing branch/rate applies. - **Unavailable** — no applicable published/reference pricing exists for the requested configuration/time. - **Invalid** — the submitted scenario violates validation or availability rules.

CTXDEX must not present incomplete, ambiguous, unavailable, or invalid results as an ordinary complete total.

Capacity bill vs effective workload economics

Help

CTXDEX keeps the canonical capacity charge separate from derived cost-per-workload analytics.

For provisioned or dedicated configurations, the Provider may bill capacity in units such as throughput-unit-hours or dedicated deployment time.

That **canonical capacity bill** remains authoritative. If you also supply workload/utilization assumptions, CTXDEX can derive analytical economics such as effective cost per million tokens or per request.

These derived workload economics are scenario-dependent comparison metrics; they are not rewritten Provider Rates and should be labelled accordingly.

About this estimate

Help

The Calculator result is an estimate derived from CTXDEX pricing data and your scenario; it is not a Provider invoice.

CTXDEX calculates an **estimate** from the selected commercial configuration, pricing-effective time, scenario usage, applicable meters, Rates, Services, and other pricing conditions.

The result is intended for planning and explanation. It is not observed Provider telemetry, a reconciled invoice, or a settled billing record.

The charge breakdown and Calculation Explanation show how the estimate was derived and which material assumptions affected it.

Industry comparison

Model and offering pages say which external products each of ours sits beside. This is where those claims come from — how a peer was chosen, what the strength words are used for, and every external model the comparisons draw on.

How a peer is chosen

Peers per subject
Two to five, three intended
Weak peers
Not used as filler
Price in choosing a peer
Counts for nothing
Method
CTXDEX_COMPARABLES_V1
Captured as of
2026-09-14

The strength words, and how often each is used

These are the data partner’s words and the comparison contract does not define them. Rather than put our sentence under their label, this says how often each is used on each side. A model can be compared closely while the commercial surface it is sold on is only a reference — that is why the two columns disagree.

Reference

35 model comparisons · 292 offering comparisons

Close

95 model comparisons · 132 offering comparisons

Direct

44 model comparisons · 65 offering comparisons

The 85 external models every comparison draws on

Unlike everything else on this site, these name real products. They are recorded as references, not as prices: where one is no longer current it stays here as history rather than being removed, and a model referenced by nothing is one that was compared against before and is not now.

ModelProviderLifecycleCheckedReferenced by
Qwen3-Coder-30B-A3BAlibaba CloudProvider lifecycle active2026-09-141 ctxdex model
Qwen3-Coder-480B-A35BAlibaba CloudProvider lifecycle active2026-09-141 ctxdex model
Qwen3-Coder-NextAlibaba CloudProvider lifecycle active2026-09-143 ctxdex models
Qwen3.5-35B-A3BAlibaba CloudProvider lifecycle active2026-09-146 ctxdex models
Qwen3.5-397B-A17BAlibaba CloudProvider lifecycle active2026-09-141 ctxdex model
Qwen3.7 Text EmbeddingAlibaba CloudProvider lifecycle active2026-09-143 ctxdex models
Qwen3-Embedding-0.6BAlibaba/QwenProvider lifecycle active2026-09-141 ctxdex model
Qwen3-Embedding-8BAlibaba/QwenProvider lifecycle active2026-09-142 ctxdex models
Qwen3-Reranker-8BAlibaba/QwenProvider lifecycle active2026-09-143 ctxdex models
Claude Opus 5AnthropicProvider lifecycle active2026-09-141 ctxdex model
Claude Sonnet 5AnthropicProvider lifecycle active2026-09-143 ctxdex models
Evo 2Arc InstituteProvider lifecycle active2026-09-142 ctxdex models
FLUX.2Black Forest LabsProvider lifecycle active2026-09-141 ctxdex model
Boltz-2Boltz/NVIDIAProvider lifecycle active2026-09-142 ctxdex models
Chai-1Chai DiscoveryProvider lifecycle active2026-09-141 ctxdex model
Cohere Embed 4CohereProvider lifecycle active2026-09-141 ctxdex model
Cohere Rerank 4 FastCohereProvider lifecycle active2026-09-142 ctxdex models
Cohere Rerank 4 ProCohereProvider lifecycle active2026-09-141 ctxdex model
DeepSeek V4.1-FlashDeepSeekProvider lifecycle active2026-09-143 ctxdex models
ECMWF AIFS ENS v2ECMWFProvider lifecycle active2026-09-141 ctxdex model
ECMWF AIFS Single v2ECMWFProvider lifecycle active2026-09-141 ctxdex model
Eleven Music v2.5ElevenLabsProvider lifecycle active2026-09-143 ctxdex models
Eleven v3 ConversationalElevenLabsProvider lifecycle active2026-09-141 ctxdex model
Scribe v2ElevenLabsProvider lifecycle active2026-09-141 ctxdex model
AlphaEarth FoundationsGoogleProvider lifecycle active2026-09-141 ctxdex model
Gemini 3 Pro ImageGoogleProvider lifecycle active2026-09-141 ctxdex model
Gemini 3.1 Flash ImageGoogleProvider lifecycle active2026-09-142 ctxdex models
Gemini 3.1 Flash LiveGoogleProvider lifecycle preview2026-09-143 ctxdex models
Gemini 3.1 Flash-LiteGoogleProvider lifecycle active2026-09-141 ctxdex model
Gemini 3.1 ProGoogleProvider lifecycle active2026-09-142 ctxdex models
Gemini 3.5 TranscribeGoogleProvider lifecycle active2026-09-14No current model
Gemini 3.8 FlashGoogleProvider lifecycle active2026-09-144 ctxdex models
Gemini Embedding 2GoogleProvider lifecycle active2026-09-141 ctxdex model
Gemini Omni FlashGoogleProvider lifecycle active2026-09-142 ctxdex models
Gemma 4 12BGoogleProvider lifecycle active2026-09-143 ctxdex models
Gemma 4 31BGoogleProvider lifecycle active2026-09-145 ctxdex models
Gemma 4 E4BGoogleProvider lifecycle active2026-09-141 ctxdex model
Lyria 3.5GoogleProvider lifecycle active2026-09-143 ctxdex models
MedGemma 1.5 4BGoogleProvider lifecycle active2026-09-142 ctxdex models
MedGemma 27BGoogleProvider lifecycle active2026-09-142 ctxdex models
ShieldGemma 2GoogleProvider lifecycle active2026-09-143 ctxdex models
TxGemma 27B PredictGoogleProvider lifecycle active2026-09-141 ctxdex model
TxGemma 9B PredictGoogleProvider lifecycle active2026-09-141 ctxdex model
Veo 3.1GoogleProvider lifecycle active2026-09-141 ctxdex model
Veo 3.1 FastGoogleProvider lifecycle active2026-09-141 ctxdex model
WeatherNext 2GoogleProvider lifecycle active2026-09-142 ctxdex models
AlphaFold 3Google DeepMindProvider lifecycle active2026-09-141 ctxdex model
AlphaGenomeGoogle DeepMindProvider lifecycle active2026-09-142 ctxdex models
Gemini Robotics 2Google DeepMindProvider lifecycle active2026-09-143 ctxdex models
Gemini Robotics ER 2Google DeepMindProvider lifecycle active2026-09-142 ctxdex models
Gemini Robotics On-Device 2Google DeepMindProvider lifecycle limited access2026-09-141 ctxdex model
Granite DoclingIBMProvider lifecycle active2026-09-143 ctxdex models
Llama Guard 4MetaProvider lifecycle active2026-09-143 ctxdex models
Aurora 1.5MicrosoftProvider lifecycle active2026-09-143 ctxdex models
CodestralMistral AIProvider lifecycle active2026-09-141 ctxdex model
Ministral 3 14BMistral AIProvider lifecycle active2026-09-144 ctxdex models
Ministral 3 8BMistral AIProvider lifecycle active2026-09-143 ctxdex models
Mistral Large 3Mistral AIProvider lifecycle active2026-09-142 ctxdex models
Mistral Medium 3.5Mistral AIProvider lifecycle active2026-09-143 ctxdex models
Mistral Moderation 2Mistral AIProvider lifecycle active2026-09-142 ctxdex models
Mistral OCR 4.1Mistral AIProvider lifecycle active2026-09-143 ctxdex models
Mistral Small 4Mistral AIProvider lifecycle active2026-09-149 ctxdex models
Voxtral Mini Transcribe 2Mistral AIProvider lifecycle active2026-09-141 ctxdex model
Prithvi-EO-2.0NASA/IBMProvider lifecycle active2026-09-141 ctxdex model
GR00T N1.6NVIDIAProvider lifecycle active2026-09-143 ctxdex models
GPT-5.6 LunaOpenAIProvider lifecycle active2026-09-141 ctxdex model
GPT-5.6 SolOpenAIProvider lifecycle active2026-09-142 ctxdex models
GPT-5.6 TerraOpenAIProvider lifecycle active2026-09-143 ctxdex models
GPT-Image-2.5 FlareOpenAIProvider lifecycle active2026-09-141 ctxdex model
GPT-Image-2.5 SunburstOpenAIProvider lifecycle active2026-09-141 ctxdex model
GPT-Live-1OpenAIProvider lifecycle active2026-09-143 ctxdex models
GPT-Realtime-2.1OpenAIProvider lifecycle active2026-09-142 ctxdex models
GPT-RosalindOpenAIProvider lifecycle limited access2026-09-144 ctxdex models
GPT-TranscribeOpenAIProvider lifecycle active2026-09-141 ctxdex model
OpenAI omni-moderationOpenAIProvider lifecycle active2026-09-141 ctxdex model
PaddleOCR-VL-1.6PaddlePaddleProvider lifecycle active2026-09-143 ctxdex models
Runway Gen-4.5RunwayProvider lifecycle active2026-09-142 ctxdex models
Stable Audio 3.0Stability AIProvider lifecycle active2026-09-143 ctxdex models
Stable Diffusion 3.5 LargeStability AIProvider lifecycle active2026-09-141 ctxdex model
Stable Image UltraStability AIProvider lifecycle active2026-09-142 ctxdex models
Voyage 4Voyage AIProvider lifecycle active2026-09-141 ctxdex model
Voyage 4 LargeVoyage AIProvider lifecycle active2026-09-142 ctxdex models
Voyage 4 LiteVoyage AIProvider lifecycle active2026-09-141 ctxdex model
Voyage Rerank 3Voyage AIProvider lifecycle active2026-09-142 ctxdex models
Voyage Rerank 3 LiteVoyage AIProvider lifecycle active2026-09-141 ctxdex model

About the data

Every model, provider, offering and rate on this site is invented. Nothing corresponds to a real product and no figure should be quoted as a market price. What is not invented is the structure.

About the data

13 notes

How the catalogue was built, and why it is invented.

About CTXDEX

Page section

CTXDEX is a structured model, commercial, pricing, comparison, and cost-estimation reference.

CTXDEX brings model information, commercial availability, pricing, industry references, and cost estimation into one structured catalogue.

The product deliberately separates **what a Model is** from **how it is made available and priced**. Technical model characteristics belong to the Model or its Snapshot; provider-specific access, serving, capacity, and commercial conditions belong to Commercial Offerings and related serving objects; pricing is resolved through the applicable commercial pricing authority.

CTXDEX is also designed as a knowledge resource. Glossary definitions explain what concepts mean, while methodology and help content explain how CTXDEX represents and uses those concepts.

What information is included

Page section

CTXDEX organizes Model, commercial, pricing, performance, comparison, and calculator information as separate but connected layers.

CTXDEX organizes information into several connected layers:

- **Model information** — identity, family, modalities, representations, capabilities, tasks, technical characteristics, openness, licence, lineage, and related model facts. - **Commercial and serving information** — publishers, inference providers, Commercial Offerings, Serving Surfaces, delivery modes, processing tiers, capacity options, commitments, services, and availability history. - **Pricing information** — published rate-card structures, meters, Rates, allowances, commitments, capacity charges, service charges, and effective-dated pricing. - **Performance information** — serving-performance reference metrics such as latency or throughput, always interpreted with their workload and serving conditions. - **Industry-reference information** — dated Similar Industry Model comparisons with explicit comparison dimensions and source references. - **Calculator information** — configured workload quantities, assumptions, estimated charge components, and calculation explanations.

Current counts, coverage, and freshness dates are derived from catalogue/read-model data rather than embedded in this editorial text.

How Model Information works

Methodology

CTXDEX keeps intrinsic Model facts separate from Snapshot overrides, Commercial Offering facts, and serving-dependent information.

CTXDEX uses semantic ownership rules so that information is stored and interpreted at the level where it actually belongs.

**Model-level information** describes the model itself: identity, family, modalities, tasks, capabilities, technical characteristics, openness, licence, and similar intrinsic facts.

**Snapshot information** records a dated or immutable model state. A resolved Snapshot view combines Model-level facts with Snapshot-specific overrides; absence of a Snapshot override does not by itself mean the Snapshot lacks that capability or characteristic.

**Commercial Offering information** describes how a Model or Snapshot is made commercially available by a Provider. Provider, access path, delivery mode, processing tier, capacity, commitment, service availability, and related commercial facts are therefore not treated as intrinsic Model properties.

**Serving performance** is normally deployment-dependent. Latency and throughput can vary with provider infrastructure, Serving Surface, processing tier, capacity, region, workload, and concurrency, so ordinary serving-performance values belong to the applicable serving/commercial context.

CTXDEX also preserves applicability states carefully. **Unknown** means a meaningful value is not currently known or verified; it does not mean unsupported or Not Applicable.

Model vs Snapshot

Methodology

A Model is the continuing model identity; a Snapshot is a dated or immutable state of that Model.

A **Model** is the canonical identity that continues across its recorded lifecycle. A **Model Snapshot** is a specific dated or immutable state linked to that Model.

CTXDEX resolves a Snapshot by combining the Model-level state with Snapshot-specific overrides:

```text Model-level facts + Snapshot-specific overrides = resolved Snapshot view ```

This matters because a Snapshot does not need to repeat every Model fact. If a Snapshot does not override a Model-level capability or representation, that absence must not automatically be interpreted as lack of support.

Snapshot release dates, identifiers, lifecycle facts, and genuine Snapshot-specific changes remain available where recorded.

Publisher vs Provider

Methodology

The Publisher releases the Model; the Provider makes it available through a serving or commercial channel.

CTXDEX keeps **Publisher** and **Provider** as separate roles.

A **Publisher** is the organization responsible for releasing or publishing a Model as part of its model portfolio or technical lineage.

A **Provider** is the organization that makes the Model available through an inference or commercial serving channel.

The same organization can perform both roles, but the distinction remains important. A Model can be published by one organization and served by multiple Providers, and those Providers can offer different access paths, capacity options, performance characteristics, and prices.

How Pricing works

Methodology

CTXDEX represents pricing as commercial, effective-dated data rather than assigning one price directly to a Model.

CTXDEX does **not** assign one naked price to a Model. Price is commercial: the same Model can be offered by different Providers, through different commercial paths, under different pricing conditions.

For model inference, pricing normally starts from the selected **Commercial Offering** and resolves the applicable effective pricing authority, rate-card branch, meters, Rates, and commercial rules. Separately priced Services, routes, or customer agreements can contribute their own charge components when they are applicable.

CTXDEX preserves **provider-native billing units and denominators**. Text may be priced by input and output tokens, while other offerings can use images, seconds, minutes, pages, requests, robot-hours, throughput-unit-hours, or other provider-defined units. CTXDEX does not invent token conversions merely to make unlike prices look uniform.

Metering and monetary valuation remain separate: a Metering Profile derives the billable quantity, and a Rate values that quantity. Historical and scheduled pricing are preserved through effective dating rather than overwriting prior commercial truth.

Pricing reference vs Pricing Calculator

Methodology

The Pricing experience shows reference rate structures; the Calculator applies those structures to a configured workload.

The **Pricing** experience and the **Pricing Calculator** serve different purposes.

```text Pricing reference → shows published/reference rate-card structures and commercial pricing context

Pricing Calculator → applies canonical pricing to a configured workload and returns an estimate ```

The Calculator accepts scenario quantities and relevant commercial choices, resolves the authoritative pricing server-side, meters the scenario into billable quantities, and calculates separate charge components before producing a total.

The Calculator does not become a second pricing authority. It must re-resolve canonical pricing rather than trusting rates passed through UI navigation or hard-coded frontend logic. Its result is an **estimate**, not a provider invoice or settled billing record.

How Commercial Offerings work

Methodology

A Commercial Offering represents a Provider-specific way a Model or Snapshot is made available.

A **Commercial Offering** is the provider-specific commercial availability object around a Model or Model Snapshot.

Different Offerings of the same Model can vary in Provider, access path, availability period, Serving Surfaces, delivery modes, processing tiers, capacity options, commitments, services, and pricing. These differences do not change the intrinsic identity of the underlying Model.

A **Serving Surface** is a concrete serving interface or path within an Offering. It can carry delivery-mode, tier, capacity, or performance context, but it is not automatically a separate priced subject unless the surface itself is independently sold or priced.

CTXDEX therefore resolves availability and pricing through the applicable Commercial Offering rather than inferring commercial truth directly from a bare Model.

Similar Industry Models

Methodology

Industry comparisons are dated, dimension-specific reference comparisons and are not claims of equivalence.

CTXDEX can show **Similar Industry Models** to give users recognizable real-world reference points for synthetic ctxdex models.

These comparisons are **directional and dimension-specific**. A comparable can be selected because of reasoning, coding, multimodal capability, image generation, realtime interaction, scientific use, robotics, document intelligence, retrieval, or other controlled comparison dimensions. The recorded Comparison Basis and similarity summary explain what part of the model positioning is being compared.

A listed model is therefore **similar or broadly comparable**, not equivalent or interchangeable. CTXDEX does not assign a numerical similarity score in V1.

Industry references are externally maintained and time-sensitive. Each comparison set carries a visible captured-as-of date, supporting rationale, methodology version where applicable, and source references so users can interpret the comparison in its proper market context.

Data freshness and date semantics

Methodology

CTXDEX separates when information was captured from when the underlying fact is effective.

CTXDEX uses different dates for different kinds of truth and does not collapse them into one ambiguous “last updated” value.

- **Captured As Of** tells you when CTXDEX captured, verified, or treated a time-sensitive reference fact as its known reference. - **Effective From** and **Effective Until** describe the interval during which a commercial, pricing, lifecycle, or other effective-dated fact applies. - **Released** describes when a Model or Snapshot became available or was released. - **Verified At**, where present, records an editorial or source-verification timestamp for externally maintained reference data.

A fact can be captured before it becomes effective, or captured recently while describing an older effective period. For that reason, database `updated_at` timestamps are not used as substitutes for user-facing captured/effective dates.

Current dates shown by the UI are derived dynamically from the relevant catalogue or read-model authority.

Historical information

Methodology

CTXDEX preserves historical Models, Offerings, Snapshots, and pricing so past states can be interpreted without presenting them as current.

CTXDEX preserves historical records where they are useful for understanding model and commercial evolution.

Historical Models, Snapshots, Commercial Offerings, replacements, and pricing can remain available for reference even after they are no longer current. Effective-dated pricing is retained so an estimate for an earlier period can resolve the price that applied at that time instead of reusing the current rate.

Historical information must remain visibly historical. A deprecated or historical Offering must not be presented as current availability, and an expired Rate must not be silently carried forward as current pricing.

Likewise, a newer item is not automatically treated as the replacement for an older one unless an explicit lineage or replacement relationship is recorded.

Synthetic data disclosure

Disclosure

CTXDEX V1 uses a constructed synthetic model and pricing universe to exercise the catalogue and calculator architecture.

The CTXDEX V1 catalogue uses **synthetic ctxdex Models, Commercial Offerings, and pricing data** constructed to exercise the model-information, commercial, pricing, comparison, and calculator systems.

Synthetic ctxdex Models and prices must **not** be interpreted as actual external vendor products, published external rate cards, or claims about real-world provider commercial terms.

CTXDEX also contains separately maintained **real-world industry reference data**, such as Similar Industry Models. Those records are externally researched, source-referenced, and visibly dated; they are not part of the synthetic model universe.

Where serving-performance data is marked with a **Synthetic Profile** Measurement Basis, it is reference/demo performance for the stated workload and serving conditions, not measured external-provider telemetry or a contractual SLA.

Known limitations

Disclosure

CTXDEX is a structured reference and estimation system; some facts are unavailable, context-dependent, synthetic, or qualitative.

CTXDEX should be interpreted with the following limitations in mind:

- Some technical characteristics can be unavailable, undisclosed, or not applicable. **Unknown** does not mean zero, false, unsupported, or Not Applicable. - Serving performance can vary materially by Provider, Serving Surface, region, processing tier, capacity, concurrency, and workload. Performance values should be read with their workload profile, Measurement Basis, statistic, and captured-as-of context. - Similar Industry Model comparisons are qualitative, dimension-specific references. They are not benchmark equivalence, interchangeability, or a numerical similarity score. - Publicly represented pricing does not necessarily describe negotiated enterprise terms. Customer-specific Commercial Agreements can have different commercial treatment. - Cost Calculator results are estimates derived from the selected scenario and applicable pricing; they are not provider invoices or observed settled charges. - Historical and externally changing information should be interpreted using the visible effective and captured-as-of dates. - Synthetic CTXDEX models, pricing, and synthetic-profile performance values are constructed reference data and must not be presented as external vendor facts.

Words this glossary does not define

These appear in prose on other pages and have no entry here yet. They are listed so the gap is visible rather than discovered.

  • Deprecation — used on Model Detail's lifecycle section, and the journey
  • Retirement — used on Model Detail's lifecycle section, and the journey