contextjet-ai/awesome-llm-observability/estimate-llm-cost

Estimate LLM cost

Cost surprises come from not doing the arithmetic. This skill ships a small, dependency-free calculator so you can price a call, project a monthly bill, and compare models with actual numbers.

Use the bundled script

`scripts/llm_cost.py` is pure Python, no install needed:

from llm_cost import estimate_cost, project_monthly
estimate_cost(1500, 300, model="gpt-6.1-sol")                       # one call, USD
project_monthly(1500, 300, calls_per_day=5000, model="gpt-6.1-sol") # monthly projection
estimate_cost(2000, 200, model="claude-sonnet-5-5", cached_input_tokens=1800)  # with prompt caching

Run it directly to see a worked example: python scripts/llm_cost.py.

The PRICES table holds the current flagship / mid / small model from OpenAI, Anthropic and Google at standard-tier short-context rates, dated by PRICES_AS_OF; two Gemini entries are promotional and step up on the date noted in the file. Prices change often, so pass input_price/output_price explicitly when you need exact figures, or edit the table. The default cached_discount=0.9 matches OpenAI; Anthropic and Google cache reads are closer to 95% off, so pass cached_discount=0.95 there. The arithmetic (not the price table) is what the tests pin down.

How to apply it

1. Price the call with realistic token counts (measure them from a trace, see instrument-llm-observability).
2. Project the bill with your real call volume. A cheap call at 10k/day beats an expensive one at 10/day.
3. Compare models by running the same tokens through two model ids. Pick the cheapest that still passes your evals (see compare-llm-models).
4. Feed it into cost cutting (see reduce-llm-cost).

Validation

Run the tests: pytest skills/estimate-llm-cost/tests/. They check the math is exact for explicit prices, that model lookups match the table, that prompt-cache discounting is correct, and that monthly projection scales linearly.

Anti-patterns

• Trusting the built-in price table as current (verify against provider pricing).
• Pricing one call and forgetting to multiply by real volume.
• Comparing models on price without checking quality holds on your eval set.