AI Agent Cost Calculator: per Run, Month and Successful Task

For each role independently, fixed prefix $F=S+T$ contains the system prompt and tool schemas. With fresh suffix $q$, visible output $o$, tool result $v$, reasoning $r$ and retained fraction $f$, history grows by $g=q+o+v+fr$. Request $t$ contains $I_t=F+q+(t-1)g$. With caching off, total input is $n(F+q)+g n(n-1)/2$: the quadratic history term described by exe.dev (2026-02-03).

The stable prefix before the current fresh suffix is $L_t=I_t-q$. If it meets the provider minimum, it is cached at the configured token boundary. A live previous checkpoint supplies read tokens; the newly extended portion is written. A request-start gap at or above the TTL conservatively expires the checkpoint, so the full eligible prefix is written again. The current fresh suffix and any rounding remainder use ordinary input pricing. Reads refresh the checkpoint lifetime. The ledger enforces $I_t=U_t+H_t+W_t$ and charges $(U_tP_i+H_tP_h+W_tP_w+(o+r)P_o)/10^6$. Write rates replace ordinary input rates for those tokens.

Anthropic's explicit prefix method, 5-minute / 1-hour lifetimes, write premiums, minimum lengths and read rates follow its pricing and prompt-caching pages, checked 2026-10-27 (s1, s12). The legacy GPT-4.1 estimate uses the published input/read/output prices (s2) with assumed retention and approximate cache boundaries; read the validity limits. Standard and batch charges are blended separately for each token category. Tool fees are per 1,000 invocations and receive no batch discount.

Expected branch charge is probability × copies × one independent branch. API/infrastructure plus review gives complete-attempt cost $C$. For success probability $p>0$ and at most $k$ independent attempts, delivery probability is $Q=1-(1-p)^k$, and expected attempts stopping at first success are $A=Q/p$. Without a cap, expected attempts to first success are $1/p$. Including one-off triage $D$, cost per success is $(AC+D)/Q=C/p+D/Q$. Monthly spend is tasks started × $(AC+D)$; expected successes are tasks started × $Q$.

The separate reliability output $p^k$ means all k trials succeed, in the sense of τ-bench passᵏ. It is not the retry delivery probability. For heterogeneous tasks, τ-bench averages per-task probabilities; raising a pooled success rate to k is only a homogeneous approximation.

• One turn is one billable model request. Each role has constant per-turn token counts and append-only history. Fixed prompt and tools remain identical. A cache breakpoint precedes the current fresh suffix; previous fresh content joins stable history on the next call.
• Each attempt, role and branch copy starts cold. There is no cross-task or cross-model cache sharing, eviction before TTL, prefix mutation, compaction or partial carry-over between retries. Thinking carried into history is explicitly controlled; actual thinking retention can be more complicated.
• The branch's success effect is already included in the supplied end-to-end success probability. Branch copies are additive cost, not an assumed accuracy improvement. Information returned to the worker belongs in its tool-result tokens.
• Attempts are independent, equally costly, and have the same success probability. A reliable evaluator detects success before retrying. Human review is charged on every attempt; one-off triage is charged on every started task.
• All money is USD, before tax. Currency values are numeric accounting amounts with USD in labels, not physical units. Time inputs have physical units and convert through hours for review. Counts and probabilities are dimensionless.
• Prices are dated examples, checked 2026-10-27. They are editable snapshots rather than live quotes. The default branch prompt remains below Haiku 4.5's cache minimum and therefore pays regular input rates.

Use for text-token API budgeting with 1–200 calls per role, positive success probability, and no more than 200,000 combined input and output tokens per request. The 200,000-token ceiling and 32,768-token output ceiling are conservative calculator scope limits, not claims about every model's full context window. Long-context surcharges, images, audio, storage, taxes, regional premiums, reserved capacity and subscription allowances are outside the model; enter separately billed infrastructure as an explicit fee.

Cache lifetime is measured from request start, including generation time. At exactly the stated TTL this model assumes expiry. Anthropic cache availability also requires valid breakpoints, unchanged settings and reachable history blocks. It assumes the immediately preceding checkpoint remains discoverable. Lower-than-minimum prefixes are not written.

GPT-4.1 is a deprecated legacy example. Its short TTL is a conservative 5-minute assumption; its long TTL is an optimistic 24-hour upper bound, not a guarantee. A 1,024-token minimum and 128-token rounding approximate the published behavior, but hidden system context, model-dependent checkpoint spacing, request settings, routing and early eviction can reduce actual hits. Edit these fields or choose no caching for a conservative estimate. Newer OpenAI models have different write/retention policies; do not apply this row to them unchanged.

Batch shares are only for eligible asynchronous API calls, not an automatic discount on sequential interactive work. Model the actual start intervals. Anthropic batch discounts stack with caching. The GPT-4.1 row conservatively leaves its cache-read price undiscounted because the published batch table gives no cached-input rate. Tool charges remain full price, and retrieved search content still needs token billing.

Success is a measured input, not a prediction. Correlated errors, harder retried cases, variable failure costs, evaluator mistakes and rate limits can invalidate the independent-attempt estimate. Zero success makes cost per success undefined and is excluded. Cache savings may be negative when writes do not earn enough reuse. HAL (arXiv:2510.11977) motivates evaluating cost and task quality together. Actual log import remains future work.