Frontier Model Cost Benchmark

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

Simulating exact token usage across probabilistic agent loops
Cost-Optimal: GPT-6 Luna ($0.038 / task)
Best Effective Cost / Completed Task
$0.038
GPT-6 Luna (72% cheaper than Legacy 5.6)
Effective Task Success Probability
98.8%
Avg 1.28 attempts with 3 max retries
Projected Monthly Production Cost
$950
Saves $2,450/mo vs Commodity Baseline
Effective Cost per Resolved Task (Micro-Breakdown) Input tokens + Reasoning + Failed Retry Waste
Direct Execution Tokens
Reasoning / CoT Overhead
Failed Attempt Waste Cost
Model Architecture Nominal $/M Pass@1 Avg Attempts Completion % Cost / Completed Task Monthly Impact
Ready. Adjust inputs to model effective cost transitions.

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.

Enjoy this tool? Build your own with Super