LAST UPDATED · JUNE 2026
ATOM Inference Price Index (AIPI) Methodology
Investment-grade standards for AI pricing intelligence
A reference guide to the construction, composition, and standards behind ATOM's inference price benchmarks. Deterministic indexing, transparent construction, rigorous quality controls. Built by information economists with 25+ years advising C-suite teams on pricing strategy.
PART ONE
Methodology Standards
Why AIPI meets institutional investment-grade benchmarks
AIPI methodology was developed by pricing intelligence specialists with 25+ years advising enterprise C-suite teams on pricing strategy. It follows financial index methodology standards comparable to S&P, MSCI, and Bloomberg indexes.
§ 1.1
Why investment-grade
AIPI meets institutional investment standards through six pillars that ensure transparency, reproducibility, and accuracy.
DETERMINISTIC
Rules-based logic, not AI estimation or web scraping.
REPRODUCIBLE
Same inputs always produce same outputs.
COMPOSITION-ADJ
Chained matched-model isolates real price movements.
TRANSPARENT
Publicly documented construction process.
AUDIT TRAIL
All data points tracked from source to index.
ANALYST OVERLAY
Human review for edge cases and validation.
§ 1.2
Data quality controls
Every price in AIPI undergoes rigorous quality controls to ensure institutional-grade accuracy.
STANDARD
IMPLEMENTATION
Deterministic extraction
Every price is extracted using rules-based logic programmed to recognize vendor-specific pricing page formats. No AI interpretation, no estimation, no guesswork.
Zero estimation
We never interpolate, extrapolate, or estimate pricing. If a vendor does not publish a price, it is not included in the index.
Analyst verification
Analysts manually review new vendor pricing page formats, ambiguous pricing structures, significant price movements, and regional availability.
Audit trail
Original and normalized prices are stored with extraction timestamps, enabling full verification of any index value back to source data.
Data freshness
Automated monitoring detects vendor pricing updates and processes changes within one business day of publication.
§ 1.3
Historical data depth
AIPI historical data is available from December 2025 forward. For vendors onboarded after this date, pricing history begins at the vendor's onboarding date. Historical data is updated retroactively when vendors publish prior period pricing information.
PERIOD
COVERAGE
December 2025 – Present
Full coverage for all vendors onboarded by December 2025
New vendors
History begins at onboarding date; retroactive backfill if vendor provides historical pricing
Index calculation
Indexes are calculated for all periods with sufficient vendor representation
PART TWO
Index Composition
16 benchmarks across modality, channel, tier, and licensing
June 2026 update. The family was broadened to sixteen indexes with the addition of AIPI EMB GLB, a global text embedding benchmark. The channel, tier, and licensing indexes now span the full token universe, text and multimodal, where earlier definitions measured text alone. The broadened definition is applied across the entire history, including the year-to-date anchors, so the series stays internally consistent throughout.
§ 2.1
Index categories
AIPI organizes 16 benchmarks across four orthogonal dimensions, each answering a distinct buyer question.
MODALITY
7 INDEXES
What does this type of inference cost? Text, image, audio, video, voice, multimodal, embedding.
CHANNEL
4 INDEXES
Where should I buy inference? Direct API, marketplaces, platforms, neoclouds.
TIER
4 INDEXES
What is the premium for flagship intelligence? Frontier, mid-tier, budget, reasoning.
LICENSING
1 INDEX
How much cheaper is open-source inference across all channels?
§ 2.2
Modality indexes
All modality indexes are calculated globally across all tracked vendors, regardless of vendor origin or region.
INDEX
UNIT
COVERAGE
AIPI TXT GLB
PER 1K TOKENS
Text generation models from leading AI vendors
AIPI MML GLB
PER 1K TOKENS
Multimodal models from leading AI vendors
AIPI IMG GLB
PER IMAGE
Image generation models from leading AI vendors
AIPI AUD GLB
PER MINUTE
Audio transcription and generation models
AIPI VID GLB
PER SECOND
Video generation models from leading AI vendors
AIPI VOC GLB
PER 1K CHARS
Voice and speech synthesis models
AIPI EMB GLB
PER 1K TOKENS
Text embedding models from leading AI vendors
§ 2.3
Channel indexes
Channel indexes track token-priced models, text and multimodal, per 1,000 tokens across four distinct distribution channels globally, enabling direct comparison of comparable models across different points of sale.
INDEX
CHANNEL
COVERAGE
AIPI DEV GLB
DIRECT API
Models priced directly by developers (OpenAI, Anthropic, Google, Mistral)
AIPI CLD GLB
CLOUD MARKETPLACE
Models priced through cloud marketplaces (AWS Bedrock, Vertex, Azure)
AIPI PLT GLB
INFERENCE PLATFORM
Models priced through inference platforms (DeepInfra, Fireworks, Together AI)
AIPI NCL GLB
NEOCLOUD
Models priced through neoclouds (Groq, Cerebras, Cloudflare Workers AI)
§ 2.4
Tier indexes
Tier indexes track token-priced models, text and multimodal, per 1,000 tokens segmented by model capability level globally.
INDEX
TIER
COVERAGE
AIPI FTR GLB
FLAGSHIP
Frontier flagship models representing peak capability
AIPI MID GLB
PRODUCTION
Mid-tier workhorse models balancing capability and cost
AIPI BDG GLB
ECONOMY
Budget economy models optimized for cost efficiency
AIPI RSN GLB
REASONING
Chain-of-thought reasoning models from leading AI vendors
§ 2.5
Licensing index
The licensing index tracks open-source and open-weight models, text and multimodal, per 1,000 tokens across all distribution channels globally, measuring the cost advantage of open-source inference relative to the broader market.
INDEX
UNIT
COVERAGE
AIPI OSS GLB
PER 1K TOKENS
Open-source and open-weight models across all channels
§ 2.6
Data sources
METHOD
DESCRIPTION
Direct extraction
Automated extraction from official vendor pricing pages using deterministic rules-based logic
API ingestion
Direct ingestion via vendor APIs where available, ensuring real-time accuracy
Analyst curation
Manual review of new vendor formats, ambiguous structures, and significant price movements
PART THREE
Index Construction
From SKU qualification to chained index value
§ 3.1
Inclusion criteria
To be included in AIPI, a SKU must meet all of the following requirements.
CRITERION
REQUIREMENT
Publicly listed
Price must be published on the vendor's website
Pay as you go
Standard on-demand pricing only
Production ready
Generally available models only
Matching unit
Price must be convertible to the index's normalized unit
§ 3.2
Exclusions
The following pricing types are excluded from AIPI.
PRICING TYPE
REASON
Negotiated rates
Not publicly verifiable
Enterprise contracts
Not publicly verifiable
Committed use discounts
Not comparable to on-demand pricing
Volume tiers
Only base tier (lowest volume) is indexed
Batch pricing
Not comparable to real-time inference
Subscriptions
Not comparable to usage-based pricing
Bundled models
Cannot attribute price to single modality
Free tiers and trials
Not representative of production costs
Beta and preview
Pricing may change at general availability
Legacy models
No longer actively offered
§ 3.3
Normalization
Vendors publish prices in different formats, units, and currencies. AIPI normalizes all prices to enable direct comparison.
VENDOR FORMAT
NORMALIZED TO
CONVERSION
per 1M tokens
USD per 1K tokens
÷ 1,000
per 1K tokens
USD per 1K tokens
No change
per image
USD per image
No change
per million pixels
USD per image
× (resolution ÷ 1M) at 1080p
per second (video)
USD per second
No change
per minute (audio)
USD per minute
No change
per second (audio)
USD per minute
× 60
per character
USD per 1K characters
× 1,000
Currency conversion: Most tracked vendors price natively in USD. For CNY pricing, ATOM uses a fixed annual exchange rate (set January 1st, held for 12 months) to eliminate FX volatility from price trend analysis. For other non-USD currencies, daily spot rates are used. Both original and normalized prices are stored for audit purposes.
§ 3.4
Calculation method: chained matched-model
AIPI uses a chained matched-model methodology to isolate actual vendor price movements from changes in tracked model composition. This approach is standard in financial index construction (comparable to CPI matched-model methodology) and prevents model additions or removals from creating false price signals. The 5-step pipeline runs every period:
01
Base period
Earliest period sets initial level using simple unweighted average
02
Match
Identify SKUs present in both current and prior period
03
Measure
Compute % change of matched-set average
04
Chain
Apply % change to prior period's index level
05
Absorb
New SKUs enter the matched set the following period
Index_t = Index_(t-1) × (avg_price_matched_t / avg_price_matched_(t-1))
Equal weighting: Within each matched set, all vendors and models are weighted equally. This prevents market dominance by high-volume providers from skewing the index and ensures pricing trends from emerging vendors remain visible alongside established players. Each index is calculated separately for the pricing directions that apply to it. Most token indexes carry Input, Cached Input, and Output; embedding is Input only, and several media indexes carry a subset.
Benchmark series: The chained value this pipeline produces is what AIPI reports as the benchmark price. It is the composition-adjusted series that movement is read from, whether week-over-week, monthly, or year-to-date. The next section sets it alongside the spot price and explains how the cached direction is measured.
§ 3.5
Benchmark price, spot price, and the caching cohort
AIPI reports two complementary price levels for each index and measures the cached direction over the models that actually offer it. The three concepts answer different questions and are shown side by side.
BENCHMARK PRICE
The composition-adjusted, chained series. It removes the effect of which models enter or leave the basket, so movement reflects real vendor repricing rather than changes in coverage. This is the value the index reports for trend reading.
SPOT PRICE
The prevailing average across qualifying SKUs live in the period, composition included. It answers a different question: what a buyer actually meets in the market right now, before any matched-model adjustment.
CACHING COHORT
Cached input is reported over the models that publish a cached rate, measured with their own input and output. Pricing it this way keeps the cached figure honest instead of diluting it against models that never offered caching.
Benchmark and spot move together over time but diverge within any single period, and that gap carries information. Benchmark shows where the matched market is trending, while spot reflects the level actually on offer in the period. Reported together, with the cached cohort measured on its own terms, they give a fuller read than any single number alone.
§ 3.6
Methodology refinements
Two safeguards strengthen the chained methodology against edge cases that can affect any matched-model index in a fast-moving market.
SAFEGUARD
PURPOSE
Per-period change bound
Matched-set percentage change is bounded each period. Swings beyond the boundary are recorded but capped, which prevents a single outlier SKU in a sparse matched set from compounding errors across the chain.
Raw YTD anchor
Year-to-date change is computed against the raw simple-median of underlying SKUs as of the first period of the current year, not against the chained value. This anchors YTD to a real observable market value.
§ 3.7
Update schedule
ACTIVITY
FREE TIERS
PAID TIERS (TERMINAL, MCP PRO, FEED)
Price extraction
Weekly
Intra-week
Index calculation
Weekly
Intra-week
Dashboard refresh
Not available
Instant
Price change alerts
Not available
Real-time
New vendor onboarding
As available
As available
Tier index review
Quarterly
Quarterly
Data freshness: Every SKU carries a Last Verified Date confirming when pricing was last checked against the vendor's published page. Automated pipelines detect vendor updates and refresh the database promptly.
PART FOUR
Reference
Glossary and scope
§ 4.1
Glossary
Standardized AIPI terminology ensuring consistency across indexes, reports, and exports.
TERM
DEFINITION
SKU
A unique price point for a specific model, vendor, and direction combination.
Index
A benchmark calculated from composition-adjusted average prices of SKUs matching specific criteria.
Matched set
SKUs present in both current and prior period, used to calculate period-over-period change without composition bias.
Chained index
Index value derived by applying matched-set percentage change to the prior period's level.
Benchmark price
The composition-adjusted, chained index value. The series ATOM reports for trend reading.
Spot price
The prevailing average across qualifying SKUs live in a period, composition included. The level a buyer meets in the market today.
Caching cohort
The models that publish a cached input rate. The cached figure is measured over this cohort using their own input and output.
Normalized price
A vendor's price converted to AIPI's standard unit and currency for comparability.
Direction
Whether a price applies to Input, Cached Input, or Output tokens.
Channel
Distribution pathway: Model Developer, Cloud Marketplace, Inference Platform, or Neocloud.
Neocloud
AI-specialized cloud provider offering serverless per-token inference on custom or purpose-built silicon.
Tier
Capability classification: Frontier (flagship), Mid-tier (production), Budget (economy), or Reasoning.
WoW
Week-over-week percentage change from the prior period's index level.
MoM
Month-over-month percentage change from four periods prior.
YTD
Year-to-date change measured against the raw simple-median anchor at the start of the current year.
§ 4.2
Scope
AIPI is deliberately optimized for investment-grade pricing intelligence. The following defines what AIPI is designed to deliver and what falls outside its scope.
OPTIMIZES FOR
Price transparency
Clear, normalized pricing data enabling direct vendor comparison
Trend detection
Identifying pricing movements, competitive dynamics, and market shifts
Vendor positioning
Understanding competitive positioning through pricing strategy
Budget forecasting
Reliable data for AI infrastructure cost planning and modeling
OUT OF SCOPE
Performance / latency
AIPI tracks cost, not speed; performance benchmarks require separate infrastructure
Quality scores
AIPI tracks specifications but does not score output quality; MMLU and HumanEval require separate testing
Total cost of ownership
Excludes infrastructure, egress, support, and integration costs which vary by deployment
Fine-tuning / training
Only inference is tracked; training costs follow different pricing structures
