← Agent Frameworks 🕐 12 min read

Agent Frameworks

FOCUS Standard and FinOps Foundation AI Working Group: LLM Token Cost Tracking for Enterprise AWS Bedrock

Executive Answer · March 2026

FOCUS 1.1–1.4 has **no native LLM-specific fields** — no token count columns, no model name columns, no inference-type columns.

Research date: 2026-06-16 Pillar: 21-benchmarks / enterprise AI FinOps standards Status: Field-verified on FOCUS 1.1 columns; docs-verified on FOCUS 1.2–1.4 and AWS Bedrock CUR; inferred on FOCUS 1.5 roadmap; community-pattern on tooling.

FOCUS 1.1–1.4 has no native LLM-specific fields — no token count columns, no model name columns, no inference-type columns. Token-level Bedrock cost lives in the AWS-proprietary x_UsageType and x_Operation columns that AWS appends to its FOCUS exports. The standard FOCUS columns (ServiceCategory, ServiceName, ConsumedQuantity) provide coarse AI cost visibility but cannot distinguish input tokens from output tokens, cached reads, or cross-region routing without AWS’s proprietary extensions.

FOCUS 1.5 (planned, no confirmed date — estimated mid-2027 by the Tokenomicon conference schedule) is explicitly scoped to add AI model identity and token consumption (input/output) to the Cost and Usage dataset. The Linux Foundation’s Tokenomics Foundation, announced June 3, 2026, will develop the companion standards and benchmarks. Until 1.5 ships, every FOCUS-based AI cost approach relies on vendor-specific columns or out-of-band log joins.

For AWS Bedrock specifically: Use CUR 2.0 (not legacy CUR) with the FOCUS 1.2 AWS export. Join x_UsageType (encodes region + model + token type + tier + routing) with x_Operation (now granular: InvokeModelInference / InvokeModelStreamingInference since January 2026) and iamPrincipal/* tag columns (CUR 2.0 only). For per-request attribution, join CUR to Bedrock model invocation logs — CUR does not carry a per-request ID.

FOCUS Specification Version Map

Version	Ratified	Core Column Count	AI/ML Changes
FOCUS 1.0	July 2024	43	None — no AI-specific columns
FOCUS 1.1	~Nov 2024	50	Added `ServiceSubcategory` (includes “Generative AI” as a valid value), `SKUMeter`, `SKUPriceDetails` — still no token fields
FOCUS 1.2	2025	~52	AWS ships this as `FOCUS_1_2_AWS`; adds `TIME_GRANULARITY` config (HOURLY/DAILY/MONTHLY); no token fields
FOCUS 1.3	Dec 5, 2025	~60+	Contract Commitment dataset; shared cost splitting columns; data recency/completeness flags — no AI token fields
FOCUS 1.4	June 4, 2026	~100+ (47 new cols)	Invoice Detail and Billing Period datasets for invoice reconciliation — no AI token fields; FOCUS 1.5 explicitly reserved for AI token columns
FOCUS 1.5	Planned ~2027	TBD	Scoped: AI model identity + token consumption (input/output) in Cost and Usage dataset; may integrate with OpenTelemetry GenAI semantic conventions

Source: FOCUS 1.4 announcement, FOCUS spec home

FOCUS 1.1 Column Reference for AI/ML Workloads

All 50 FOCUS 1.1 columns are general-purpose cloud billing fields. The columns most relevant to AI/ML workloads:

Column	Type	Feature Level	AI/ML Relevance
`ServiceCategory`	String	Mandatory	Enumerated value “AI and Machine Learning” applicable to Bedrock
`ServiceSubcategory`	String	Recommended	Values include “Generative AI”, “Machine Learning”, “Natural Language Processing”, “AI Platforms”
`ServiceName`	String	Mandatory	“Amazon Bedrock” — identifies the service
`ResourceType`	String	Conditional	Identifies resource kind; for Bedrock this is sparsely populated
`ConsumedQuantity`	Decimal	Conditional	Volume of metered usage — maps to token counts in theory but is not labeled as such
`ConsumedUnit`	String	Conditional	Would hold “tokens” if populated by provider — AWS uses `x_UsageType` instead
`PricingQuantity`	Decimal	Mandatory	Volume used for pricing calculations
`PricingUnit`	String	Mandatory	Unit of pricing — could be “tokens” but AWS encodes this in `x_UsageType`
`SKUPriceId`	String	Conditional	Price identifier — maps to a specific token rate for a model tier
`ChargeCategory`	String	Mandatory	“Usage” for Bedrock inference; “Purchase” for Provisioned Throughput commitments
`Tags`	String	Conditional	Key-value metadata — critical for team/project allocation

Gap: FOCUS 1.1 has no field for model name, token direction (input vs output vs cache read vs cache write), inference tier (standard/priority/flex), or cross-region routing. These live in AWS vendor extensions.

AWS FOCUS Exports: What Bedrock Looks Like

FOCUS 1.0 AWS (table: `FOCUS_1_0_AWS`) — 5 AWS proprietary columns

Column	Description
`x_UsageType`	Encodes region, model, token type, service tier, routing. This is the primary Bedrock cost-attribution field.
`x_Operation`	Since Jan 2026: `InvokeModelInference` or `InvokeModelStreamingInference` (was generic “Usage” before)
`x_ServiceCode`	AWS service code (e.g., `AmazonBedrock`)
`x_CostCategories`	AWS Cost Categories groupings
`x_Discounts`	Discount breakdown

FOCUS 1.2 AWS (table: `FOCUS_1_2_AWS`) — 3 AWS proprietary columns (reduced)

AWS trimmed the proprietary columns in 1.2: x_Discounts, x_Operation, x_ServiceCode. The x_UsageType and x_CostCategories columns dropped out of FOCUS 1.2 exports — this is a breaking change for Bedrock cost slicing if teams migrated from 1.0 to 1.2 without accounting for this.

FOCUS 1.2 adds TIME_GRANULARITY configuration (HOURLY/DAILY/MONTHLY) which is significant for Bedrock workload analysis.

Source: AWS FOCUS 1.0 with AWS columns docs, AWS FOCUS 1.2 with AWS columns docs

Bedrock CUR 2.0 (legacy and FOCUS) Usage Type Format

{region}-{model}-{token-type}                         # standard tier, in-region
{region}-{model}-{token-type}-{tier}                  # priority or flex tier
{region}-{model}-mantle-{token-type}-standard         # bedrock-mantle requests
{region}-{model}-{token-type}-cross-region-global     # cross-region inference

Four token types billed separately:

Token Type	Usage Pattern	Cost Direction
Input tokens	`-input-tokens` or `-mantle-input-tokens-*`	Lower unit cost
Output tokens	`-output-tokens` or `-mantle-output-tokens-*`	4–5x input cost
Cache read tokens	`*-cache-read-input-token-count`	Significantly cheaper than input
Cache write tokens	`*-cache-write-input-token-count`	More expensive than input

Example x_UsageType values:

USE1-Claude4.6Sonnet-input-tokens
USE1-Claude4.6Sonnet-cache-read-input-token-count
USE1-Claude4.6Sonnet-output-tokens-cross-region-global
USE1-Nova2.0Lite-input-tokens-flex
USE1-gpt-oss-120b-output-tokens-priority

Critical reconciliation warning (from AWS docs): CUR does not carry per-request IDs. Both legacy CUR and CUR 2.0 aggregate Bedrock costs by usage type, operation, and pricing/resource over an hour or a day. To attribute costs to individual requests or prompts, join CUR to Bedrock model invocation logs at model + usage-type grain.

IAM principal attribution (available only in CUR 2.0 / AWS Data Exports): tag columns appear as iamPrincipal/{key} — requires CUR 2.0 and tag activation in the Billing console (24-hour lag).

Source: AWS Bedrock CUR documentation, January 2026 granular operation visibility announcement

FinOps Foundation AI Working Group

Working Group Identity

Name: FinOps for AI Working Group (part of the FinOps Foundation, not a separate entity)
Governance: Operates under FinOps Foundation framework at finops.org/topic/finops-for-ai/
Companion: Tokenomics Foundation (announced June 3, 2026 at FinOps X; Linux Foundation-hosted; founding members: Oracle, Google, Microsoft, Accenture, Booking.com, Flexera, IBM, JPMorganChase, KPMG, Nebius, Salesforce, SAP, ServiceNow)

Published Papers and Guides (2025–2026)

Title	Published	Key Content
GenAI FinOps: How Token Pricing Really Works	Aug 1, 2025	Token pricing mechanics; hidden costs in GenAI services; cache token accounting
Scaling Kubernetes for AI/ML Workloads with FinOps to Optimize Value	Oct 3, 2025	GPU cluster FinOps; elastic operations
Navigating GenAI Capacity Options	Feb 4, 2026	Capacity models across AWS/GCP/Azure; provisioned vs on-demand tradeoffs
FinOps for AI: Tools & Services Considerations	Apr 23, 2026	AI tooling costs across training, tuning, inference, orchestration, operations lifecycle
Effect of Optimization on AI Forecasting	2025	Four-step cost estimation process; 99% token reduction case study via caching/hashing
AI Cost Estimation	2024–2025	Fundamentals of cost estimation for AI services
How to Forecast AI	2024–2025	Forecasting techniques for AI workloads
Optimizing GenAI Usage: A FinOps Perspective	TBD	Prompt engineering, caching, model selection optimization
Choosing an AI Approach and Infrastructure Strategy	TBD	Infrastructure strategy for AI deployments

Source: FinOps for AI topic page, FinOps Framework AI technology category

FinOps Framework AI Capability Model

The FinOps Framework organizes AI cost management across four domains:

1. Understand Usage & Cost

Data Ingestion, Allocation, Reporting & Analytics, Anomaly Management
Key challenge identified: “more acute difficulties identifying AI model consumers” when multiple interfaces share resources

2. Quantify Business Value

Planning & Estimating, Forecasting, Budgeting, KPI & Benchmarking, Unit Economics
Recommended KPIs: Cost Per Token, Token Consumption Efficiency, Cost Per Inference
Forecast volatility warning: require “more frequent revision” and “shorter forecasting windows” during early adoption

3. Optimize Usage and Cost

Architecting & Workload Placement, Usage Optimization, Rate Optimization
Specific tools recommended: RouteLLM, vLLM, LLMLingua, Tiktoken, SkyPilot

4. Manage the FinOps Practice

Risk/Policy/Governance, Automation/Tools, Intersecting Disciplines
Recommended governance structure: AI Investment Council (Finance + Product + Engineering + Procurement + Leadership)

Four-Step AI Cost Estimation Framework (from “Effect of Optimization” paper)

Identify application archetype
Estimate volume of AI services consumed (tokens, API calls, images)
Populate application component inventory
Analyze total application cost

Token Economics Definition (FinOps Foundation)

Token economics = the discipline governing “how energy and capital are converted into AI tokens, how those tokens are consumed efficiently, and how that spend connects to business value.”

Key metrics:

Cost per inference and per outcome
Token yield rate — generated tokens contributing to business actions post-retries
Token consumption efficiency — operational cost relative to token usage
Revenue or value per megawatt (for GPU operators)

Practical FinOps Tooling — FOCUS and Bedrock Coverage

Tool Comparison Matrix

Tool	FOCUS Native	Bedrock Support	Token-Level Tracking	Notes
Vantage	Yes — FOCUS-native reports; custom provider CSV must follow FOCUS schema	Yes — CloudWatch S3 logs integration (private preview)	Yes — cost per 1K tokens, per model, per team	Only tool with MCP server; LLM Token Allocation in private preview for OpenAI + Bedrock
CloudHealth (Broadcom)	FOCUS 1.0 integrated	Yes — AI cost observability; practitioner playbooks	Partial — observability-layer; less granular than Vantage	Strategic roadmap includes FOCUS standardization and AI spend management
IBM Apptio / Cloudability	FOCUS-aware (FOCUS data spec referenced in pipeline)	Via AI Investment TCO product (GA May 2025)	Yes — AI Investment TCO provides full AI cost transparency; Cloudability Governance integrates with Terraform	New Conversational Insights feature (preview, June 2026)
Kubecost (acquired by Apptio)	Yes — FOCUS export	Kubernetes workloads only; does not track Bedrock API calls	No — tracks infrastructure cost of pods running ML code, not token API spend	Kubecost 3.0 launched with Apptio Cloudability Governance integration
OpenCost	Yes — opencost-plugins accept FOCUS-schema custom cost sources	No dedicated Bedrock plugin	No — plugins directory contains only: Datadog, MongoDB Atlas, OpenAI (verified via GitHub API, June 2026)	2026 roadmap lists “AI usage cost tracking” as priority; no implementation shipped
Pay-i	Not confirmed	Not confirmed	Yes — purpose-built for AI FinOps; cost per user/model/token dashboards	Emerging specialized tool
nOps	Partial	Anthropic Claude integration (via AWS)	Limited	AWS-focused; AWS Bedrock costs surfaced through AWS cost data
Spot by NetApp / Granulate	Not confirmed for FOCUS	No specific LLM/Bedrock coverage found	No	Focus is GPU infrastructure optimization, not API token spend
Anodot	Not confirmed	Not confirmed	Anomaly detection for cost spikes — applicable to Bedrock	No specific FOCUS or token-level Bedrock docs found
LiteLLM	N/A (not a FinOps tool — it’s an LLM proxy)	Yes — native Bedrock support	Yes — token cost tracking across 100+ LLM APIs including Bedrock, configurable callback to cost tracking systems	Open source; commonly used as middleware layer that feeds cost data to FinOps tools

Source: Vantage LLM Token Allocation, Best FinOps tools for AI, Apptio AI Investment TCO announcement

OpenCost Plugin Architecture (docs-verified)

OpenCost plugins require conformance to the FOCUS spec for custom cost source ingestion. Plugin structure:

opencost/opencost-plugins/pkg/plugins/
├── datadog/
├── mongodb-atlas/
└── openai/        # No Bedrock equivalent exists as of June 2026

The plugin contract: gather cost data for a time window and resolution, emit in FOCUS-compatible schema. A Bedrock plugin could be built following the OpenAI plugin as a template, consuming Bedrock model invocation logs from CloudWatch/S3.

Source: opencost/opencost-plugins GitHub — verified via GitHub API June 2026.

Architecture: How Bedrock Token Cost Attribution Actually Works

Bedrock API call
    │
    ├─► Bedrock model invocation logs (CloudWatch/S3)
    │       per-requestId, model, input_tokens, output_tokens, latency
    │       ← ONLY source with per-request granularity
    │
    └─► CUR 2.0 / AWS Data Exports (hourly/daily aggregate)
            line_item_usage_type: USE1-Claude4.6Sonnet-input-tokens
            line_item_operation: InvokeModelInference (since Jan 2026)
            iamPrincipal/tag-key: {caller identity} (CUR 2.0 only)
            resourceTags/key: {project/team tags}
            ← FOCUS 1.2 AWS export adds TIME_GRANULARITY config
            ← x_UsageType in FOCUS 1.0 AWS; NOT in FOCUS 1.2 AWS

JOIN at: model × usage_type × hour/day grain
    → reconciles dollars (CUR) with request volume and metadata (logs)
    → enables: cost per team, per project, per inference profile

FOCUS layer (what gets standardized):
    ChargeCategory = "Usage"
    ServiceCategory = "AI and Machine Learning"
    ServiceSubcategory = "Generative AI"
    ServiceName = "Amazon Bedrock"
    ConsumedUnit = (not standardized by AWS in FOCUS — lives in x_UsageType)
    Tags = team/project/IAM principal tags

What FOCUS 1.5 promises to add (~2027):
    ModelId (or equivalent) = model family identifier
    InputTokenCount = token volume for pricing
    OutputTokenCount = token volume for pricing
    InferenceType = on-demand / provisioned / cross-region

Hard Edges and Known Gaps

Gap	Impact	Mitigation
FOCUS 1.0–1.4 has no token count columns	Cannot cross-provider-compare AI costs in FOCUS without vendor extensions	Use `x_UsageType` (AWS) + model invocation logs; wait for FOCUS 1.5
`x_UsageType` absent in FOCUS 1.2 AWS export	Teams migrating from FOCUS 1.0 to 1.2 AWS lose the key Bedrock slicing column	Stay on FOCUS 1.0 AWS for Bedrock analysis until AWS adds equivalent to 1.2
CUR has no per-request ID	Cannot attribute cost to specific prompts from CUR alone	Enable Bedrock model invocation logging; join logs to CUR at model+hour grain
Cache read and cache write are separate line items	Reconciliation fails if only input+output tokens are summed	Account for all 4 token types in any cost calculation
IAM principal attribution requires CUR 2.0	Legacy CUR users cannot do caller-level attribution	Migrate to CUR 2.0 / AWS Data Exports
OpenCost has no Bedrock plugin	OpenCost cannot natively track Bedrock API token spend	Build plugin following OpenAI template; or use Vantage/LiteLLM as sidecar
FOCUS 1.5 has no confirmed GA date	AI token columns in FOCUS are roadmap, not shipping	Enterprise governance on AWS relies on x_UsageType + invocation logs until 2027+
Tokenomics Foundation governance charter not published	Standards for cross-provider AI billing normalization are not yet defined	Monitor Tokenomicon Amsterdam (Sep 22–23, 2026) for first working drafts
Multi-agent workload cost allocation	No accepted framework for allocating costs across multi-agent pipelines that share a model	FinOps Foundation working group identifies this as active gap; no solution published

Verification Ledger

Claim	Evidence Level	Source	Date
FOCUS 1.1 has 50 columns, no AI token fields	Source/docs verified	Fetched FOCUS 1.1 spec page directly	June 2026
FOCUS 1.4 ratified June 4, 2026; 47 new columns; no AI fields	Docs verified	finops.org/insights/introducing-focus-1-4/	June 2026
FOCUS 1.5 scoped for AI model identity + token consumption	Docs verified (one source)	Amnic guide + efficientlyconnected.com	June 2026
AWS FOCUS 1.0 has 5 x_ columns incl x_UsageType	Docs verified	AWS CUR docs fetched directly	June 2026
AWS FOCUS 1.2 has only 3 x_ columns (x_UsageType dropped)	Docs verified	AWS FOCUS 1.2 docs fetched directly	June 2026
Bedrock CUR usage type format (4 token types, tier encoding)	Docs verified	AWS Bedrock CUR docs fetched directly	June 2026
Bedrock x_Operation now shows InvokeModelInference since Jan 2026	Docs verified	AWS whats-new page fetched directly	June 2026
Tokenomics Foundation announced June 3, 2026; founding members listed	Docs verified	Linux Foundation press release, FinOps X keynote	June 2026
OpenCost plugins: Datadog, MongoDB Atlas, OpenAI only	Source verified	GitHub API query on opencost/opencost-plugins	June 2026
Vantage LLM Token Allocation in private preview for Bedrock	Docs verified	Vantage blog post fetched	June 2026
Apptio AI Investment TCO GA May 2025	Docs verified	Apptio press release	2025
FinOps for AI working group paper list	Docs verified	FinOps Foundation topic page fetched	June 2026

Next Research Lanes

Tokenomicon Amsterdam (Sep 22–23, 2026) — first concrete output from Tokenomics Foundation; watch for working draft of AI billing schema that may precede FOCUS 1.5.
AWS FOCUS 1.3/1.4 AWS export — AWS is currently on FOCUS 1.2 AWS. Check whether AWS ships a FOCUS 1.3 or 1.4 variant and whether x_UsageType returns.
FOCUS for AI dataset proposal — the FinOps Foundation is actively weighing “a net-new observability dataset vs. integration with OpenTelemetry GenAI semantic conventions.” Monitoring the FOCUS GitHub spec repo for an AI dataset RFC would be high-value signal.
Vantage LLM Token Allocation GA — currently private preview. GA announcement would confirm whether FOCUS-schema CSV upload is the data ingestion model for Bedrock token data.
LiteLLM as a FinOps middleware layer — LiteLLM tracks token costs across 100+ APIs including Bedrock. An integration between LiteLLM’s cost callbacks and an OpenCost Bedrock plugin would close the gap between API-layer token data and Kubernetes infrastructure cost.
Multi-agent cost allocation framework — the FinOps Foundation has identified this as an unsolved problem. Watch for a working group paper in late 2026.