Per-Task AI Economics Lab
Nominal token pricing is deceptive. When models require verification passes, reasoning tokens, and agent retry loops, a "cheaper" token price can easily result in 4x higher cost per resolved task. Architect your real unit economics below.
Economics & Efficiency Analysis
| Model Architecture | Nominal $/M | Pass@1 | Avg Attempts | Completion % | Cost / Completed Task | Monthly Impact |
|---|
Why Per-Task Pricing Dominates Per-Token Pricing
When Sam Altman and AI labs discuss moving toward per-task pricing, they reflect a mathematical reality of autonomous systems: nominal token cost only measures compute volume, not problem-solving efficiency.
If a cheap model costs $0.20 per million tokens but only possesses a 32% Pass@1 rate on complex tool use, an autonomous agent will fail repeatedly, consume massive scratchpad tokens, trigger validation retries, or leave tasks unresolved. A frontier model at $1.50 per million with an 85% success rate routinely solves the task in a single shot—resulting in less total dollars spent per accomplished task.
Mathematical Framework of this Engine
Expected Attempts & Probability of Resolution
Given single-turn pass rate $p$ and max retries $K$, the probability of task completion is $P_{\text{success}} = 1 - (1 - p)^K$. Expected attempts $\mathbb{E}[N] = \frac{1 - (1 - p)^K}{p}$.
Token Accumulation & Failed Retry Waste
Every failed attempt incurs full context ingestion, tool schemas, and generation tokens without yielding production utility. Effective unit cost divides the entire expected spend by the probability of delivered success.
Reasoning Token Multiplier
Models with extended test-time compute (like GPT-6 Sol or o-series) emit internal reasoning tokens before delivering the answer. This tool explicitly prices reasoning tokens alongside visible output.