Occupant Indices

Public measurement infrastructure I built for the AI economy. Three live indices that turn scattered public data into numbers you can track over time: the price tag, the activity meter, and the labor gauge for AI work.

Occupant dashboard on desktop: a ticker strip of all eight readings, $CPI 62.4 on an orange chart block, $AIU 504.1 and $LDI 3.86% beside it, and the four tier prices below — The Occupant home screen, desktop. Every index and tier price on one surface, June 2, 2026.

The premise

Every major input to the economy has a published number you can look up: a spot price for crude, a CPI print for groceries, a wholesale rate for electricity, a capacity factor for a wind farm. Occupant builds the same kind of instrument for AI compute: a standardized, dated, methodology-driven yardstick that anyone (a procurement officer, an investor, a policy researcher, or another model) can read cold and act on.

Everyone gets the same access, the way they do to a public yield curve.

The indices

Compute Price Index screen: 62.4 on an orange block, down 9.7% in one month, with an all-time chart and tier prices — $CPI, May 16, 2026. The basket price, the chart since launch, and the four tiers.

AI Market Index screen: 504.1 on a sky blue block, up 2.2% in two weeks, with the three weighted components below — $AIU. Token volume, inferred spend and the energy proxy, weighted 60 / 30 / 10.

Labor Displacement Index screen: substitution rate 3.86% on a yellow block, with nine federal workloads listed by human and AI cost per unit — $LDI. Nine federal workloads, human cost against AI cost per unit.

$CPICompute Price Indexthe price signal

Measures: the price of AI work at the basket level.

$CPI defines a fixed basket of standardized workloads and re-prices that basket every day across capability tiers, volume-weighted over 2,188 models and normalized to February 2025 = 100. It’s a deflator, so the basket stays constant while its cost moves.

Capability tiers: $BULK (commodity models under $1/MTok), $FRONT (frontier capability, ranked by Arena ELO), $JUDGE (reasoning-intensive, o1 / o3 / R1-class), $LCTX (long-context, 128K+).
Build Cost sub-indices apply realistic, fixed workload mixes so a reader lands on the number that matches their own usage: $START (startup builder), $AGENT (agentic systems), $THRU (high-throughput).
Published across multiple horizons: since-launch, year-over-year, quarter- and month-to-date, week-over-week.

Reading it: at ~62.4, a basket of AI work costs about 38% less than at launch, and fell ~9.7% in the most recent month.

Workload tiers screen: $FRONT $11.33, $JUDGE $10.37, $LCTX $1.39, $BULK $0.60, and the all-tier basket at $25.68 per 1,000 workloads — The four capability tiers and the volume-weighted basket.

Who pays what screen: three buyer profiles, five basket methods as bars, price deflation by horizon, and the cognition premium of $10.73 — Build Cost sub-indices by buyer profile, and deflation across five horizons.

$AIUAI Economic Activity Indexthe volume signal

Measures: the volume of AI economic activity, expressed as a synthetic unit (the AIU) modeled on the IMF’s Special Drawing Rights.

The AIU is a weighted composite:

60% token throughput, for usage intensity.
30% inferred spend (tokens × blended pricing), for economic scale.
10% energy proxy: a blend of token-derived energy estimate (70%) and global grid-investment growth (30%), so the energy term carries independent information.

Baseline February 2025 = 100. At ~504 it reflects roughly a 5× expansion in activity over about fifteen months. Weekly volume is on the order of 42.7 trillion tokens and $210M of inferred spend.

Reading them together: total AI spend ≈ AIU × CPI. This decomposes the headline spending figure into its two drivers, efficiency gains (price) and demand growth (volume), which is the split the bubble debate needs.

$LDILabor Displacement Indexthe labor signal

Measures: the pressure AI puts on real jobs, reported as two distinct numbers that answer two distinct questions.

Cost differential (structural). What AI could displace at today’s pricing. Across nine pilot federal workloads, average human cost runs ~$53.56/unit against ~$0.0062/unit for AI, a cost ratio on the order of 37,000× and a structural displacement potential near $4.2B/year.
Substitution rate (observable). What procurement records show actually shifting: ~3.86%.
Absorption dimension. When a workload automates, the index classifies whether the workforce is cut, frozen, or reallocated, because “cheaper” and “fired” are separate events.

SNAP eligibility screen: $25.13 per human review against $0.0003 per AI review, a 98,176 times cost ratio, and 16,500 federal staff marked reallocated — The SNAP eligibility deep dive. Every number on the screen traces to a named source.

AI procurement screen: Department of Defense obligated $1.18B, up 12%, with nine more agencies ranked by FY2024 spend and change — AI procurement by agency, FY2024 against FY2023, from USAspending.

Built on named public data: BLS OEWS + ECEC wages for human cost, FPDS / USAspending contractor spend as substitution proxy, OPM FedScope FTE counts for headcount, mapped through SOC and PSC crosswalks.

Worked end-to-end: the SNAP Eligibility deep dive resolves all three pillars to one real program: ~$25.13 per human review vs ~$0.0003 for AI, ~41M reviews/year, a ~2.1% substitution signal, and an “absorption: reallocated” verdict as caseload grows against a flat workforce. Every step is auditable on the page.

The tools

The tools let people act on the index readings.

LLM Cost Calculator. Enter a monthly spend or raw token volumes, pick a workload profile, and compare a tiered model mix at live $CPI rates. Answers “lock in now or wait a quarter?”
LDI Workload Calculator. Run the human-vs-AI cost and substitution math on a workload of your own.
AI Services Price Reasonableness Worksheet. A print-ready, FAR-style determination sheet a contracting officer can fill in, sign, and attach to an RFP, built on the $CPI basket.
Market Intelligence. “Sabermetrics for models”: Quality-Adjusted Price, cognitive arbitrage (≈90% of frontier capability at ≈40% of the cost), and tier analysis across the market.

Market intelligence screen: weekly tokens 41.8T, reasoning share 7.4%, provider share of volume, five value leaders by quality adjusted price, and average price and Elo by tier — Market Intelligence. Quality-adjusted price puts Gemini 2.0 Flash at the top of the value table.

Model prices screen: blended price per million tokens for each model, grouped by Frontier, Reasoning and Budget tier with a bar for each — Model prices per million tokens, grouped by tier, with the tier average on each header.

How it’s built

The architecture follows directly from the infrastructure-first mandate.

Frontend: Vanilla HTML5 / CSS3 / JavaScript. Flat files on any static host.
Data pipeline: Python 3 in src/: dedicated fetchers per source (OpenRouter, LiteLLM, BLS, FPDS / USAspending, OPM FedScope, BloombergNEF grid data), calculators (calculate_cpi, calculate_aeai, calculate_ldi), a model-tier registry, signal derivation, and historical backfill.
Storage: Flat-file JSON in data/, including machine-readable endpoints another agent can pull as a grounded, dated source.
Automation: GitHub Actions on a daily cron recompute the indices and commit fresh JSON. The git history is the data lineage: every published number is reproducible from the commit log.
Algorithms: Deterministic, documented formulas with published weights and baselines.
Privacy: Client-side localStorage for the theme toggle only.
License: Parachute Commons: free for personal, educational, and non-commercial use; tiered terms at commercial scale.
Cost to operate: ~$50/month.

What it took

Economic index design. Baskets, volume weighting, normalization, base periods, and a synthetic composite unit (the AIU’s SDR analogy), applied to a new domain.
Data engineering on messy public sources. Reconciling price discrepancies across OpenRouter and LiteLLM, building SOC ↔ PSC ↔ FedScope crosswalks, and designing fallback hierarchies.
Domain range, coherently connected. CPI / macro methodology, federal procurement (FAR, FPDS, FedRAMP), labor economics (BLS, JOLTS, OPM), and energy capital expenditure.
Product and communication judgment. Designing for a thirty-second read by a procurement officer and holding a deliberate editorial voice.
Pragmatic systems architecture. Static-site-plus-cron with git as an audit log; three indices that compose, where the LDI consumes the live CPI basket.

The three indices are public, recomputed daily, and auditable.