pray per token
Lexicon ex Machina
pray per token /preɪ pɜːr ˈtoʊ.kən/ adj., n.
adjective
Describing a workflow dependent on external API services over which the user has no control, continuity guarantee, or pricing stability. “His entire product was pray per token—one rate hike from insolvency.”
Characterized by anxious dependence on third-party inference providers. “A pray per token architecture: latency spikes at 4pm, outages during demos, deprecation without warning.”
noun
The economic model in which AI capabilities are rented rather than owned, leaving practitioners spiritually and financially exposed. “Pray per token is the sharecropping of the compute era.”
The state of hoping an API call succeeds, returns something useful, and doesn’t cost more than expected. “He hit submit and entered pray per token—would this be the $0.002 call or the $4 one?”
Canonical Anxieties
Undocumented rate limits
Silent model swaps (”we’ve upgraded you to a faster model” that is worse)
Context window changes
The invoice that arrives three weeks late
“This model has been deprecated”
Derivatives
pray per tokener n. One who subsists on API calls. “A pray per tokener learns to cache aggressively and trust nothing.”
Usage Note Distinct from pay per token, the official pricing model, in that pray captures the supplicant relationship between user and provider. The pray per tokener does not negotiate; he hopes.
See Also GPU-poor, vendor lock-in, rug pull
Origin Mid-2020s, pun on pay per token. Emerged from developer communities processing the reality that their applications existed at the pleasure of three companies.