Count every model call
Include planning, execution, reflection, evaluation, and fallback calls. Agent cost is usually call volume multiplied by context growth, not one visible chat request.
AI agent cost calculator
Estimate the cost of the whole agent loop.
A single task can trigger planning, retrieval, tool use, validation, and retries. Model those calls together instead of pricing one prompt in isolation.
Free · no sign-upCalculations run in your browser.Pricing registry reviewed Jul 31, 2026
Same workload, five lowest estimates
Use this as a shortlist, then test quality and retry behavior on your own tasks.
- DeepSeek V4 FlashDeepSeek$17.75/ month
- GPT-5.6 LunaOpenAI$43.24/ month
- Llama 4 MaverickMeta$47.12/ month
- DeepSeek V4 ProDeepSeek$54.79/ month
- Gemini 3.5 Flash-LiteGoogle$79.55/ month
This models token charges, not hosting, vector storage, search tools, human review, or provider minimums. Cache savings apply only to models with a published cached-input rate.
Include retries honestly
Failed tool calls, malformed output, and quality retries still consume tokens. A small retry assumption makes forecasts more realistic and exposes fragile routes.
Compare cost per completed task
Per-call cost is useful for debugging. Cost per accepted task is better for business decisions because it includes multi-step behavior and failures.
Plain-English answers
Questions people ask before using the result
What should count as an LLM call in an agent?
Any request to a language model: planning, tool selection, summarization, execution, self-checking, fallback, and final response generation.
How do I estimate tokens per call?
Measure representative production traces when possible. Otherwise start with the assembled system prompt, conversation history, retrieved context, tool schemas, tool output, and expected response.
Does this include vector databases and tool fees?
No. It estimates model token charges. Add embeddings, vector storage, search, browser tools, hosting, and human review separately.