Skip to content

Cost and FinOpsGuide

Can FOCUS billing data carry AI token spend? A field mapping

FOCUS can hold AI spend beside cloud cost, but token billing only maps cleanly in part. This guide goes line item by line item, covering input, output, cache, batch, service tier and media. It marks where a custom column or a documented convention is still needed.

By INFRO Editorial

Published

9 min read

Glowing token streams being compressed into a precise efficient AI infrastructure pipeline

In short

Yes, in part. FOCUS 1.4 can carry AI usage in its Cost and Usage dataset, and the 1.5 release candidate adds labels for model, cache action and token direction. Batch, service tier, modality and reasoning effort still have no standard home, so plan for custom x_ columns.26

Key takeaways

  • FOCUS 1.4 can hold AI usage today as ordinary Cost and Usage rows; what it lacks is shared labels for model, cache action and token direction, which the 1.5 release candidate adds as recommended properties.
  • Map each meter (input, output, cache read, cache write) to its own row and SKU, and never fold cached tokens into input rows, or token totals double count.
  • Batch, service tier, modality and reasoning effort have no standard property in the 1.5 candidate, so carry them as documented x_ columns and keep the value sets stable.
  • Gateway cost is a rate calculation: put it in ListCost and ContractedCost, and fill BilledCost only from the provider's own reported figure.
  • Write a one-page mapping contract per provider before loading anything, and test it by summing the dataset back to the provider's invoice.

The short answer

FOCUS is described as an open specification that normalizes billing datasets across AI, cloud, SaaS and other vendors.1 In practice, a token line item already fits the shape of a Cost and Usage row: a charge, a quantity, a unit, a price and a cost. Where it breaks is the labeling. Two providers can sell the same model, meter caching differently, and name their meters differently, and FOCUS 1.4 gives you no common way to say which is which.

FOCUS 1.5 addresses part of that. Per the project's release-scope page, the release candidate was completed on 2 November, with ratification scheduled for 3 December 2026 and public announcement on 10 December.2 Today is 18 November, so 1.5 is a candidate, not a published version. This guide maps onto 1.4 columns and flags where 1.5 properties would apply.

A note on how to read the mapping. The column choices below are a proposed convention for a team building its own pipeline, not text copied from the specification. Check each against the column library (opens in a new tab) before you freeze it.

Three sources, one dataset

AI spend usually arrives from three places, and each fills a different part of a FOCUS row.

  • Direct provider usage and cost exports. Accurate and late, grouped by the provider's own project or key. They carry the invoiced figure, so they are the only honest source for BilledCost.
  • The cloud bill. Bedrock, Azure and Vertex usage is already in the cloud invoice. The FOCUS home page lists AWS, Microsoft and Google Cloud generators at v1.2 when read, so check which version your cloud export produces before you expect any 1.4 column from it.1
  • Gateway request records. Timely, and they carry your own dimensions (feature, customer, team). Their cost is a rate-table estimate, not an invoice.

The Framework treats AI as a Technology Category alongside Public Cloud and SaaS,3 so the same Allocation and Reporting capabilities apply. The capability map covers that side. This guide covers the column-level work underneath it.

The mapping, line item by line item

FOCUS 1.5 says plainly that FOCUS has carried token counts since version 1.0 and that what it adds is the labels that make them comparable.2 So the usage side is not new. The table is the working proposal.

Line itemQuantity and unitWhere the label goesFit
Input tokensUsage quantity in tokensSKU Price Details: TokenDirection = Input (1.5)Clean in 1.5; custom key in 1.4
Output tokensSame, per output meterTokenDirection = OutputClean
Reasoning tokens billed separatelyOwn meterTokenDirection = Output, per the 1.5 textClean, but you lose the reasoning split
Cache readOwn meterTokenCacheAction = ReadClean in 1.5
Cache writeOwn meterTokenCacheAction = WriteClean in 1.5
Model identityNot a quantityModelDeveloper, ModelFamily, ModelId, ModelVersionRecommended in 1.5, optional
Batch input and outputSame unitsNothing standardNo clean fit
Service tier (Standard, Flex, Priority, Reserved)Same units, different rateNothing standardNo clean fit
Image, video, audio input tokensTokens per modalityNothing standardNo clean fit
Per-image, per-second, per-character, per-hour unitsNon-token unitsUsage and pricing unit columnsFits as units, not as a shared metric
Provider-issued unitsOwn currencyPricing Currency as a consumption currency (1.5)Clean in 1.5
Gateway estimateRequest-level costListCost and ContractedCostClean, with a flag
Proposed mapping of AI line items to FOCUS Cost and Usage fields, with the 1.5 release candidate noted.

The four model properties, and the two token properties, live in SKU Price Details, which the 1.5 page describes as a Cost and Usage column carrying properties that describe what a price is for, with FOCUS defining common names so queries work across providers.2 Including them is recommended, not required, so a conformant dataset can omit them.2

Tokens and cache: one row per meter

The 1.5 text gives three rules that you can adopt now. TokenCacheAction takes Uncached, Read, Write or Other, and TokenDirection takes Input or Output.2 A price that bills several kinds of tokens on one meter omits the property that would misdescribe them.2 And a cached token counted on a Read or Write row is not counted again on another row, so totals do not double count.2

The practical consequence is that you need one row per meter, not one row per request. If your provider export reports input_tokens, cache_read_tokens and cache_write_tokens as columns on a single line, unpivot it. Then check one thing before anything else: whether the provider's input count includes cached tokens or excludes them. If it includes them, subtract the cache reads from the uncached input row, or the same token is billed twice in your dataset.

Cloud providers show why labels matter. Bedrock's page notes for Amazon Nova that cache read input tokens are priced 75% below the on-demand input token price.4 With a TokenCacheAction label, a rising bill from a falling cache hit rate can be told apart from a bill that rose because of traffic. The 1.5 page makes the same point and lists cache hit rate by model as an example query.2

Batch and service tiers: where the spec has no clean fit

Bedrock offers Standard, Flex, Priority and Reserved tiers,4 and offers select models for batch inference at a 50% lower price than on-demand.4 Both change the rate for identical tokens. In the 1.5 release scope I read, neither batch nor service tier appears among the properties; the AI items it lists are model identity, cache action, token direction and a consumption currency, plus worked examples.2

That leaves you three honest options.

  1. Separate SKUs. If the provider bills batch and standard on different meters, the SKU already differs, and your reports can group by SKU. This costs nothing but is unreadable without a lookup.
  2. A custom column. Custom Columns is a supported feature in the 1.4 specification,5 and the 1.4 introduction says generators carry native dataset columns as custom x_ columns.6 Use x_processing_mode (synchronous, batch) and x_service_tier with a closed value list you document.
  3. A key inside SKU Price Details. Workable, but those are properties of a price, and the spec reserves FOCUS-defined names for itself. Use a clearly vendor-prefixed key.

Pick option 2 for anything finance filters on. Whichever you choose, the meaning belongs in a mapping contract you version, because the next person to query the data cannot guess it.

Media: units that are not tokens

Text is the easy case. The Nova 2 Omni rows on Bedrock's page list separate input prices per million tokens for text, image, video and audio, and a separate output price for image,4 so one model produces several meters with the same unit and different modalities. The 1.5 page is explicit that modality, reasoning effort, cache retention duration and context window size are not values of the new token properties.2 Modality is therefore a gap you fill yourself, with an x_modality column.

Non-token media is a different problem. INFRO's own docs list the metering units across its endpoints as token for text, image for image generation, second of output for video, character for speech and hour of audio for transcription.7 That is a normal situation for FOCUS: the usage unit and pricing unit columns can hold images, seconds or characters. What breaks is any cross-model unit metric. "Cost per million tokens" has no equivalent that you can compare against cost per generated second, so keep unit economics per modality and do not blend them into one rate.

Provider-issued units and drawn-down commitments

Some AI usage is priced in a provider's own unit rather than in a national currency. The 1.5 candidate adds a consumption currency: a provider-issued unit such as platform units or normalized billing units, with ISO 4217 currency codes reserved for national currencies.2 Under 1.4, put the provider unit in a custom column and keep Pricing Currency to a real ISO code, so a validator does not reject the row.

For commitments, the 1.4 rule is the one to know. A covering charge (a purchase that pays for other charges) carries an effective cost of 0, and its cost flows to the covered usage as that usage is consumed.6 If you track a committed-spend agreement with an AI provider, your consumption rows carry the recognized share and the purchase row must not also count. Treating both as cost double counts the commitment.

Gateway estimates: which cost columns they may fill

A gateway records tokens and computes a cost at request time. That is useful for allocation, because the request carries your feature and customer, but it is not an invoice. FOCUS defines BilledCost as the amount invoiced by the invoice issuer, and INFRO's export documentation quotes the requirement that it must reflect invoiced amounts, not estimated or inferred values.8

INFRO's FOCUS v1.4 export follows that. Its estimate goes to ListCost and ContractedCost, which carry the same figure, priced at the customer's negotiated rate when one is on record and the vendor's published rate otherwise.8 BilledCost is filled only with cost_basis=reconciled, from the provider's own cost API, allocated to rows in proportion to their share of the estimate for that provider-day.8 Otherwise it is absent, not zeroed.8 The dataset is available from GET /v1/usage/focus as JSON or CSV; the FOCUS export page (opens in a new tab) states which columns are populated and which are deliberately empty.8

Copy the discipline even if you build your own pipeline. A report that mixes estimated and invoiced rows in one cost column cannot be reconciled afterwards. Keep a flag, such as x_cost_basis with values estimated and reconciled, on every row.

Where there is no clean fit: the short list

  • Batch and service tier. No standard property in the 1.5 candidate; use documented x_ columns.
  • Modality, reasoning effort, context window, cache retention. Named in the 1.5 text as outside the token properties.2
  • Request-level attribution. Per-request feature, customer and agent are not provider billing facts. The 1.5 candidate lets cost rows name the principal and credential behind a charge,2 but a feature or workflow tag still has to come from your application and a gateway.
  • Failed and cancelled calls. Whether a provider bills them is provider-specific, and FOCUS has no property for it. See do providers bill cancelled or failed streams.
  • Response-cache hits. A request answered from a gateway cache never reaches the provider, so it has no provider charge. Record it as zero-cost usage with a custom flag or leave it out, and decide which before the first report.

A checklist to run before you load anything

  • Write a mapping contract per provider: which export field becomes which column, and which are unmapped.
  • Unpivot cache meters into separate rows and check whether provider input counts include cached tokens.
  • Add x_processing_mode, x_service_tier, x_modality and x_cost_basis with closed value lists.
  • Put estimates in ListCost and ContractedCost only; fill BilledCost from provider cost data only.
  • Enter negotiated rates so the estimate uses them; otherwise list-rate rows drift from the invoice.
  • Sum each provider-day back to the provider's reported cost and investigate any persistent gap.
  • Record the FOCUS version and the correction style of each source, since a 1.2 cloud export and a 1.4 gateway export will not share columns.
  • Re-check the mapping once 1.5 is ratified and generators adopt it.

The reconciliation step is where the effort goes. The 1.4 post introduces Invoice Detail and the Invoice Reconciliation feature for tying usage to invoices,6 and INFRO's reconciliation compares estimated spend daily with the cost reported by OpenAI's and Anthropic's own cost APIs. Method detail is in reconciling estimated LLM spend with provider invoices.

Where to start

Do not wait for a perfect schema. Pick the provider with the largest spend, write its mapping contract, load one month, and check that the totals match the invoice. The rest of the work is repeating that per provider, with the same custom columns, so one query works across all of them.

Questions readers ask

Should I wait for FOCUS 1.5 before loading AI spend?
No. Data that conforms to 1.4 stays valid as 1.4, and the 1.5 changes are additions rather than renames or removals, so rows you load now keep working.2 Put the 1.5 property values into SKU Price Details or x_ columns today so the later change is a rename of a key, not a rebuild.

From the team that builds it

Where this shows up in INFRO

INFRO is a control plane in front of the provider accounts you already hold. The capabilities below are the parts of it this article touches, each with the documentation that defines its behaviour.

  • Usage, request and FOCUS exports

    Available with a plan

    GET /v1/usage and GET /v1/requests on any gateway key, the request log and usage series as CSV, plus a FOCUS v1.4 Cost and Usage dataset as JSON or CSV, so AI spend lands in your FinOps platform beside the rest of your cloud cost.

    Read the documentation for Usage, request and FOCUS exports
  • Usage and cost per request

    Available with a plan

    Every request records its model, latency, token or unit counts, and estimated provider cost, queryable by project and key.

    Read the documentation for Usage and cost per request
  • Negotiated provider rates

    Available with a plan

    Enter the rates your contracts actually carry, per model and optionally per provider, so estimated cost uses them before the vendor's list rate. Included on every plan.

    Read the documentation for Negotiated provider rates
  • Provider cost reconciliation

    Available with a plan

    Daily comparison of estimated spend against the cost reported by OpenAI's and Anthropic's own cost APIs, using a separate admin credential, with the likely cause of a persistent gap and a suggested rate you can apply. Included on every plan.

    Read the documentation for Provider cost reconciliation
  • Spend attribution

    Available with a plan

    Group cost by model, project, key, or an end-user id you attach to the request.

    Read the documentation for Spend attribution

INFRO accounts are self-serve. Provider usage is billed to you by your providers; INFRO charges a software subscription and takes 0% of model spend. See plans.

Sources

Numbered in order of first citation. Each was opened and checked against the claim on the date shown.

  1. 1.FOCUS: FinOps Open Cost & Usage Specification (opens in a new tab) — FinOps Foundation (FOCUS project). Accessed Nov 18, 2026.
  2. 2.FOCUS 1.5 Release Scope: Confirmed, In Review and Deferred (opens in a new tab) — FinOps Foundation (FOCUS project). Accessed Nov 18, 2026.
  3. 3.FinOps Framework Overview (opens in a new tab) — FinOps Foundation. Accessed Nov 18, 2026.
  4. 4.Amazon Bedrock pricing (opens in a new tab) — Amazon Web Services. Accessed Nov 18, 2026.
  5. 5.FOCUS Specification 1.4 (opens in a new tab) — FinOps Foundation (FOCUS project). Accessed Nov 18, 2026.
  6. 6.Introducing FOCUS 1.4: Invoice Reconciliation, Commitment Details, and Specification Maturity (opens in a new tab) — FinOps Foundation. Accessed Nov 18, 2026.
  7. 7.INFRO documentation overview (opens in a new tab) — INFRO. Accessed Nov 18, 2026.
  8. 8.FOCUS export (opens in a new tab) — INFRO. Accessed Nov 18, 2026.

About this article

Read on 2026-11-18: the FOCUS home page, the FOCUS 1.4 specification index and introduction post, the FOCUS 1.5 release-scope page, the FinOps Framework overview, the Amazon Bedrock pricing page (truncated, so only tier, batch and Nova cache and modality statements are used) and INFRO's docs overview and FOCUS export page. The individual FOCUS column pages returned only navigation, so column semantics beyond what those pages state are given as a proposed mapping, not as spec text, and no provider-specific export formats are described.

Last checked against its sources on . The next review is due by . No separate named reviewer is recorded for this article.

How articles are written, sourced and corrected is set out in the editorial standards.