AI Agent Running costs in production.

Three concrete examples, each picked to match the model it would actually run on in production. Cost = input tokens × input price + output tokens × output price. Tool-call round-trips count toward the input bill on the next turn — they're the most common cost surprise teams hit.

Snapshot from openrouter.ai and provider catalogues, May 2026. Provider-direct pricing is usually within ±5% of OpenRouter's published rate; volume discounts and prompt-cache reads can drop the effective input price by 50–90%.

Per-run cost × runs/day × 30. Log-scaled axes; expensive frontier workloads pull away fast. Most teams we work with sit between 1,000 and 10,000 runs/day per agent.

Same three workloads, side by side: AI cost vs. the loaded monthly cost of the role that would otherwise own the work. Numbers assume each profile runs at 1,000 events/day (the middle of the projection above) for 30 days.

Per-run costs come from the model prices above; per-run human-equivalents use loaded US/EU salaries for the role that owns the work today (a support analyst at $28/hr, a senior engineer at $95/hr, a research analyst at $72/hr). At 1,000 events/day for 30 days, the simple agent saves ~$56k/month against a single support analyst's hourly cost; the research agent saves over $3M/month against a research desk doing the same volume manually.

• Input tokens dominate. Most agents read more than they write — a code-review agent typically processes 5–10× more input than output. Trimming the system prompt and the retrieved context is the highest-leverage cost lever.
• Tool calls inflate input. Each tool-call round-trip re-sends the prior conversation plus the tool result on the next turn. A 12-step agent on a 4k system prompt easily reads 48k tokens before the user notices.
• Prompt caching is the cheapest fix. Anthropic, OpenAI, and Google all cache the static head of the prompt at 10–25% of the standard input rate. For long, stable system prompts this drops the effective bill 50–90%.
• Match the model to the step. Use Haiku/Flash/mini for classifiers and routing, escalate to Sonnet/GPT-5/Gemini Pro only when the eval set demands it. A two-tier router cuts spend on most production agents in half.
• Reasoning models bill on hidden tokens. Frontier reasoning modes ("thinking", "deliberation") count toward output. A single Opus 4.7 reasoning call can write 4–6k tokens you'll never see.
• Per-tenant budgets in the runtime. Hard-cap each agent at a daily $ ceiling. We set this in CI and the workflow cancels promotion if the eval suite blows past it.