Concept: Cost per Accepted Result
Cost per Accepted Result is the economic and operational metric used to evaluate AI model efficiency based on the true end-to-end expense required to deliver a verified, error-free outcome, rather than relying strictly on published per-token API prices.
The Token Price Illusion
Evaluating AI models solely on input and output token rates creates a false sense of economy. A cheap per-token model (e.g., $0.87 per million tokens) can end up costing significantly more than a premium frontier model (e.g., fable-5 or chatgpt-5-6) if the cheaper model requires:
- Longer Reasoning Traces: Generating excessive hidden or verbose reasoning tokens to navigate complex logic.
- Unnecessary Tool Calls: Churning through repeated, unproductive API or tool calls.
- High Failure & Retry Rates: Failing late in multi-step workflows, forcing execution loops to restart.
- Human Cleanup & Verification: Requiring manual human review, correction, and error fixing.
Formula & Operational Components
Benchmark Evidence
During government evaluation tests by CAISI on deepseek V4 Pro across seven standard benchmarks, DeepSeek ranged from 53% cheaper to 41% more expensive per correctly solved task compared to leading US frontier models. While DeepSeek’s raw output token price was 15x to 30x lower, token verbosity and task failure rates inverted the cost advantage on complex tasks.
Strategic Guidelines
- Bounded High-Volume Tasks: For simple extraction, classification, and reviewable text processing, low-cost models excel because automated check-and-recheck loops remain inexpensive.
- Ambiguous High-Stakes Jobs: Premium models that succeed in a single pass frequently yield a lower finished work cost by avoiding iteration loops and human intervention.