AI Agent Cost Per Task Calculator

An agentic task is not one model call but a loop: the model reasons, invokes a tool, receives the result, and is called again with the transcript so far. Because most frameworks resend the entire conversation on every turn, input tokens accumulate arithmetically. If the first call carries B tokens and each turn adds roughly the same amount, the k-th call carries about k times B, and the sum across N calls is B times N times (N+1)/2, which is quadratic in the number of turns. This model applies that as an average context growth factor of (1+N)/2 per call, then adds a tool overhead percentage covering JSON schema definitions for the tools, system instructions, and serialised tool results, all of which are billed as input on every single turn. The cost asymmetry shapes optimisation. Output tokens are priced three to five times input per token, but agents emit few output tokens per call while re-reading an ever-growing prompt, so input normally dominates the bill once N passes about five. Useful ranges: a simple reason-act-observe loop finishes in three to seven calls, planning agents run five to fifteen, and open-ended research agents can exceed thirty. The model ignores prompt caching, which can cut the repeated prefix substantially, and assumes a fixed per-call token budget, so a task dominated by one very large document read will be underestimated. Capping iterations is the standard control on worst-case spend.

Chaining language model calls into a loop became practical only once models could reliably follow instructions. The immediate ancestor is chain-of-thought prompting, described by Jason Wei and colleagues at Google in 2022, which showed that prompting a model to reason step by step improved multi-step accuracy. Later that year Shunyu Yao and co-authors at Princeton and Google published ReAct, interleaving reasoning traces with actions against an external environment such as a search API. That paper established the reason-act-observe pattern most agent frameworks still implement. Tooling followed quickly. LangChain was released as open source by Harrison Chase in October 2022, and OpenAI shipped function calling in June 2023, giving models a structured way to request tool invocations and making the token overhead of tool schemas an explicit line item in every request. Autonomous loop projects such as AutoGPT, released in March 2023, made the cost problem visible to a wide audience when unbounded iteration produced surprising bills. The pricing response arrived during 2024, when providers introduced prompt caching for repeated prefixes and discounted batch endpoints, both aimed directly at the resend-the-whole-transcript economics that make agent workloads expensive relative to single-shot calls.