CDX-301e · Module 3

Cost-Per-Task Analysis

3 min read

Cost-per-task is the fundamental unit of cloud execution economics. It consists of three components: model cost (input + output tokens), compute cost (VM time), and overhead cost (boot time, dependency installation, git operations). For a typical 10-minute task, model cost is 60-80% of the total, compute cost is 15-30%, and overhead is 5-10%. The leverage is in model selection and task duration — a task that completes in 5 minutes on GPT-4.1 costs a fraction of the same task taking 15 minutes on o3-high.

When you multiply by parallelism, cost scales linearly. Ten parallel tasks cost 10x one task. Best-of-3 on five parallel tasks costs 15x. A daily batch of 50 tasks with best-of-2 costs 100x a single task. These multipliers are obvious in isolation but easy to miss when building automation. The discipline is to calculate the steady-state cost before deploying any automated batch workflow: tasks per day * cost per task * days per month = monthly cloud spend. Compare this to the value delivered — hours saved, bugs caught, quality improvement.

# Cost-per-task breakdown (illustrative)

Model: GPT-4.1
  Input tokens:    ~50K ($0.10)
  Output tokens:   ~10K ($0.04)
  Model subtotal:  $0.14

Compute:
  VM time:         8 minutes @ $0.02/min = $0.16
  Boot overhead:   30 seconds (included in VM time)
  Compute subtotal: $0.16

Total per task:    $0.30

# Cost multipliers
Best-of-3:         $0.90 per logical task
5 parallel tasks:  $1.50 per batch
Daily batch (50):  $15.00 per day
Monthly (22 days): $330.00 per month

# Optimization levers
- Model selection:  GPT-4.1 vs o3 = 3-5x cost difference
- Task scoping:     Smaller tasks = shorter VM time
- Cache hits:       Warm pools reduce boot overhead
- Retry reduction:  Better AGENTS.md = fewer failures

Do This

Calculate monthly cost projections before deploying automated batch workflows
Use GPT-4.1 for routine tasks and reserve o3 for tasks that genuinely need reasoning depth
Track cost per task type to identify which workflows deliver ROI and which are overpriced
Optimize failure rates first — eliminating retries is the highest-leverage cost reduction

Avoid This

Deploy batch automation without a budget — uncapped parallelism creates uncapped costs
Use the most expensive model for every task because "quality matters" — match model to task complexity
Ignore boot overhead — it is a fixed cost that dominates for short tasks
Forget that best-of-N multiplies the entire cost, not just the model cost