📡

AI Agent Prompt Cost Forecaster

Estimate monthly LLM API spend from prompt length, output tokens, request volume, and cached-input percentage.

Workload

Prices are directional estimates in USD. Last updated: 2026-07-07. See notes.

Estimated monthly cost with OpenAI GPT-4o mini

Enter a workload to see the cost.

Provider / model Input $/1M Cached $/1M Output $/1M Monthly cost Cost / 1k requests
OpenAI GPT-4o Strong general reasoning $2.500 $1.250 $10.000
OpenAI GPT-4o mini Cheap routing / extraction $0.150 $0.075 $0.600
Anthropic Claude 3.5 Sonnet Long-context reasoning $3.000 $1.500 $15.000
Anthropic Claude 3 Haiku Fast, cheap classification $0.250 $0.125 $1.250
Google Gemini 1.5 Pro 1M context window $1.250 $0.625 $5.000
Google Gemini 1.5 Flash Fast, cheap retrieval $0.075 $0.037 $0.300
Groq Llama 3.1 70B Low-latency inference $0.590 $0.590 $0.790
Together AI Llama 3.1 70B Open weights at API cost $0.880 $0.880 $0.880
Cohere Command R+ RAG-optimized model $2.500 $1.250 $10.000
Local RTX 4090 Effective per-token cost incl. power & amortization $0.100 $0.100 $0.100

Cost levers

Prompt tokens / request

Cached input tokens / request

Output tokens / request

Enter a workload to see cost levers.

Frequently asked questions

What is a prompt cost forecaster?

A prompt cost forecaster estimates how much you'll spend on LLM API calls before you launch an agent. It combines your expected requests per month, the size of your system and user prompts, the average output length, and the percentage of input tokens that hit a provider cache.

How is cached input different from uncached input?

Some providers charge less for input tokens they have seen recently and keep in a prompt cache (Anthropic, OpenAI, Google). If your system prompt or retrieved context repeats across requests, a higher cache-hit rate lowers your bill.

Why does output tokens matter so much?

Output tokens are usually more expensive per token than input tokens, and long outputs multiply quickly. A coding agent or report writer that emits 1,500 tokens per request can cost 3–10x more than a short classification call.

Which providers offer prompt caching discounts?

OpenAI GPT-4o/4o-mini, Anthropic Claude 3.5 Sonnet/Haiku, and Google Gemini 1.5 Pro/Flash publish discounted cached-input rates in this calculator. Other providers treat cached and uncached input at the same price.

When is self-hosting cheaper than API calls?

Self-hosting on a local RTX 4090 or rental GPU becomes cheaper at high, steady volume when you can keep the card utilized. This calculator shows an effective per-token cost for local inference so you can compare directly against managed APIs.

Prices are directional estimates per 1M tokens in USD. Provider rates change frequently; confirm current pricing before budgeting. Local GPU cost is an effective per-token estimate including power and hardware amortization.

🚀 Get AI automation insights daily

15:00 MST. One-click unsubscribe.

Subscribe