promptcost

Token budget and cost estimator for LLM prompts.

What It Does

promptcost estimates the dollar cost of running your LLM prompts against specific models. It provides:

• Per-invocation cost estimation with input/output token breakdowns
• Multi-model comparison to find the cheapest model for your prompt
• Volume projections (daily, monthly, annual costs at a given call rate)
• Budget enforcement with pass/fail exit codes for CI
• Cost delta analysis to measure the cost impact of prompt changes
• Pipeline cost estimation for multi-stage prompt workflows

Installation

pip install promptcost-ai

Dependencies: prompttools-core-ai >= 1.0, typer >= 0.12, rich >= 13.0

CLI Commands

promptcost estimate

Estimate costs for a prompt file or directory.

# Single file estimation
promptcost estimate prompts/greeting.yaml --model gpt-4o

# Estimate with specific output token count
promptcost estimate prompts/greeting.yaml --model gpt-4o --output-tokens 500

# Estimate with volume projections
promptcost estimate prompts/greeting.yaml --model gpt-4o --project 1000/day

# Compare across multiple models
promptcost estimate prompts/greeting.yaml --compare

# Compare specific models
promptcost estimate prompts/greeting.yaml --models gpt-4o,gpt-4o-mini,claude-4-sonnet

# Estimate all files in a directory
promptcost estimate prompts/ --model gpt-4o

# JSON output
promptcost estimate prompts/greeting.yaml --model gpt-4o --format json

Options:

Option	Default	Description
--model, -m	claude-4-sonnet	Model profile for estimation
--output-tokens	auto	Override estimated output tokens
--project	none	Project at volume (e.g., 1000/day, 500/hour)
--compare	false	Compare across default model set
--models	none	Comma-separated model list for comparison
--format, -f	text	Output format: text or json

promptcost budget

Check prompt costs against a per-invocation budget ceiling. Exits with code 1 if any prompt exceeds the budget.

# Check a single file
promptcost budget prompts/greeting.yaml --limit 0.05 --model gpt-4o

# Check all files in a directory
promptcost budget prompts/ --limit 0.10 --model gpt-4o

Options:

Option	Required	Description
--limit, -l	yes	Maximum cost per invocation in USD
--model, -m	no	Model profile (default: claude-4-sonnet)
--output-tokens	no	Override output token estimate

promptcost delta

Show the cost impact of a prompt change by comparing old and new versions.

# Basic delta
promptcost delta prompts/old.yaml prompts/new.yaml --model gpt-4o

# Delta with volume projections
promptcost delta prompts/old.yaml prompts/new.yaml --model gpt-4o --volume 1000/day

promptcost models

List all available model profiles with pricing information.

promptcost models

Displays a table with model name, provider, context window, input price per million tokens, output price per million tokens, and encoding.

Output Token Estimation

When --output-tokens is not specified, promptcost estimates output tokens using a heuristic based on prompt content keywords:

Detected keywords	Estimated output	Method
json, structured, schema, format	500 tokens	heuristic
brief, short, concise, summary	300 tokens	heuristic
detailed, comprehensive, thorough, essay	2,000 tokens	heuristic
None of the above	1,000 tokens	heuristic
expected_output_tokens in prompt metadata	As specified	explicit

The estimate is capped at the model's max_output_tokens limit.

Volume Projection Formats

The --project / --volume flag accepts these formats:

Format	Meaning
1000/day	1,000 calls per day
500/hour	500 calls per hour (converted to per-day)
200/week	200 calls per week (converted to per-day)
5000/month	5,000 calls per month (converted to per-day)

Projections are calculated as: daily, monthly (30 days), and annual (365 days).

Programmatic Usage

from prompttools_core import parse_file, parse_pipeline
from promptcost import (
    estimate_file,
    estimate_pipeline,
    compare_models,
    project_cost,
    check_budget,
)

# Single file estimation
prompt = parse_file("prompts/greeting.yaml")
estimate = estimate_file(prompt, model="gpt-4o")
print(f"Input tokens: {estimate.input_tokens}")
print(f"Est. output tokens: {estimate.estimated_output_tokens}")
print(f"Total cost: ${estimate.total_cost:.4f}")

# Model comparison
comparison = compare_models(prompt, ["gpt-4o", "gpt-4o-mini", "claude-4-sonnet"])
print(f"Cheapest: {comparison.cheapest}")
print(f"Savings: ${comparison.savings_vs_most_expensive:.4f}/call")

# Volume projection
projection = project_cost(estimate, "1000/day")
print(f"Monthly cost: ${projection.monthly_cost:.2f}")

# Budget enforcement
estimates = [estimate_file(pf, "gpt-4o") for pf in parse_directory("prompts/")]
results = check_budget(estimates, budget=0.05)
for r in results:
    print(f"{r.file_path.name}: {'OVER' if r.over_budget else 'OK'}")

# Pipeline estimation
pipeline = parse_pipeline("pipeline.yaml")
pipeline_est = estimate_pipeline(pipeline, model="gpt-4o")
print(f"Pipeline total: ${pipeline_est.total_cost:.4f}")

Data Models

CostEstimate -- Per-invocation cost for a single prompt.

• file_path, model, input_tokens, estimated_output_tokens
• output_estimation_method ("explicit", "heuristic", "max")
• input_cost, output_cost, total_cost

CostComparison -- Comparison across multiple models.

• estimates: dict[str, CostEstimate], cheapest, most_expensive
• savings_vs_most_expensive

CostProjection -- Volume-based cost projections.

• volume, calls_per_day, daily_cost, monthly_cost, annual_cost

BudgetResult -- Budget check result.

• file_path, model, estimated_cost, budget, over_budget, overage

PipelineCostEstimate / StageCostEstimate -- Pipeline-level cost breakdown.

CI Integration

GitHub Actions -- Budget Gate

- name: Check prompt budget
  run: promptcost budget prompts/ --limit 0.10 --model gpt-4o

GitHub Actions -- Cost Delta on PR

- name: Check cost impact
  run: |
    git show HEAD~1:prompts/main.yaml > /tmp/old.yaml
    promptcost delta /tmp/old.yaml prompts/main.yaml --model gpt-4o --volume 1000/day

Exit codes:

Code	Meaning
0	All prompts within budget / estimation succeeded
1	One or more prompts exceed budget
2	Path not found or other error

License

MIT License. Author: Scott Converse.