Agent Workload Cost Simulator

Agents call the model many times per task, and each step usually re-sends everything that came before. That is why agent bills surprise people. Model your workload below and compare what it costs per run, per day, and per month across models.

Steps per agent run (model calls)

Base input tokens per step

Avg output tokens per step

Tool-call overhead tokens per step

Agent runs per day

Context accumulates across steps Each step's input also carries every previous step's output (the normal agent loop). Turn off to model a fixed-size context per step.

Input tokens / run

Output tokens / run

Total tokens / run

Tokens / day

Cost comparison (30-day month)

Model	In $/1M	Out $/1M	Cost / run	Cost / day	Cost / month

Pricing and caching notes: Claude rows use the published Anthropic API prices per 1M tokens (as of July 2026): Claude Fable 5 $10 in / $50 out, Claude Opus 5 and Opus 4.8 $5 in / $25 out, Claude Sonnet 5 and Sonnet 4.6 $3 in / $15 out, Claude Haiku 4.5 $1 in / $5 out. Rows marked * (GPT, Gemini) are approximate - check the provider pricing page. These are list prices with no discounts: real agent workloads are the single best case for prompt caching, because every step re-sends the same system prompt, tool definitions, and conversation prefix. With caching (roughly 90% off cache reads on the Claude API) and batch processing, real bills often come in 50-90% below the numbers shown here. See the LLM Price Calculator for plain per-request pricing.