Metered consumption, pay-what-you-use models, and the metering infrastructure that makes it work.
Three schools of AI agent pricing are at war. Token-based, task-based, and outcome-based models all have fatal flaws — and everyone is shipping anyway. Here's how to pick your poison.
Compute credits drove 150%+ NRR and the biggest software IPO of its time. Then customers optimized. The 2023 guidance cut is the most important UBP lesson in a decade.
Off-by-one errors that cost real money. Clock skew across distributed systems. Idempotency failures that double-bill your best customers. What production-grade metering actually requires.
Usage-based billing (UBB) charges customers based on actual consumption — API calls, compute hours, data processed, or transactions — rather than a flat subscription fee. Companies like Snowflake, Twilio, and AWS pioneered this model, and it's now the fastest-growing pricing pattern in SaaS and AI.
A production metering pipeline needs event ingestion (often via Kafka or a streaming queue), deduplication with idempotency keys, aggregation windows, and a rating engine that applies pricing rules. The biggest risk is off-by-one errors — metering bugs cost real money on every invoice.
They're often used interchangeably, but consumption billing typically implies pure pay-as-you-go with no subscription floor (like Snowflake), while usage-based billing can include hybrid models with a base subscription plus metered overages.
Revenue unpredictability is the biggest challenge — usage can drop when customers optimize. CFOs dislike variable costs, which is why many companies add committed-spend floors or credit packages to smooth revenue while keeping usage-based expansion.
Metronome, Orb, and Lago are purpose-built for usage-based billing. Stripe Billing supports metered billing but hits limitations at scale. Zuora and Chargebee added usage support but started as subscription-first platforms.