Cost per task
An end-to-end cost perspective based on completing a single task, combining the number of calls, retry overhead, context expansion and human review time.
Core Facts
Timeline (last 90 days)
The true cost of a task should also count the number of calls, the overhead of failed retries, the input bloat caused by context length, and the time spent manually reviewing weak-model outputs.
Evaluating an agent model cannot rely on per-token price alone, because a single task typically involves dozens of model calls and tool round-trips, and a model with many retry rounds and slow convergence may have higher end-to-end total costs even if its unit price is low.
Evaluating new models should focus on three hard metrics: independent benchmark scores, token consumption per task, and response speed.
Deploying top-tier flagship models for simple tasks wastes computing resources and drives up the cost per task.
All Facts (4)
Evaluating new models should focus on three hard metrics: independent benchmark scores, token consumption per task, and response speed.
65%UnverifiedThe true cost of a task should also count the number of calls, the overhead of failed retries, the input bloat caused by context length, and the time spent manually reviewing weak-model outputs.
50%UnverifiedEvaluating an agent model cannot rely on per-token price alone, because a single task typically involves dozens of model calls and tool round-trips, and a model with many retry rounds and slow convergence may have higher end-to-end total costs even if its unit price is low.
50%UnverifiedDeploying top-tier flagship models for simple tasks wastes computing resources and drives up the cost per task.
50%