WEDGE · subsidy ×
–×
How much cheaper is a token on a flat plan than metered at API rates?
The economics of the AI subscription: a flat plan gives you AI labor at ~0 marginal cost until the cap, while a large org pays metered API rates. AISUBS quantifies that wedge — the "149-person company" subsidy — from live OpenRouter API prices, and tracks the metering clock as the all-you-can-eat era closes.
The aggregate subsidy multiple over time against the 20–70× envelope SemiAnalysis measured empirically. It drifts with live API prices — only Anthropic, on Opus-heavy traffic, reaches the envelope.
The essay hedged — "maybe the labs will meter everyone." The answer arrived: token/usage metering went near-universal in 2026 H1. The wedge is a subsidy being actively withdrawn.
Anthropic Enterprise meters every token with zero included usage; OpenAI bills Codex at API rates; Microsoft and GitHub Copilot flipped whole plans to credits. Google meters via Vertex.
The API price is a traffic-weighted blend across each provider's models. Benchmark against Opus-tier and the wedge runs toward 70×; against Flash it's near parity. This is the live mix that sets each blend.
For each flagship coding plan, the API-rate value of the tokens you can burn at max usage ÷ the plan price. SemiAnalysis measured this empirically at 20–70×; we reconstruct it live and band the one modeled input.
Anthropic's Team plan caps at 150 seats, then forces Enterprise — $20/seat + metered API, zero included. The only clean public token-metering notch. CLIFF measures the per-seat jump at that line.
METER tracks the share of major providers whose big-org path already meters tokens. Honest caveat: the API-price history varies usage mix, not each model's past list price — true forward prices accrue from launch.
One number for the subsidy a flat plan delivers — and the clock on how long it lasts.