Cost per Run and the AI Cost Stack

The unit of AI spend

Cost per run is the dollar cost of one execution of a workflow. Monthly run cost is that number multiplied by how many times the workflow runs per month. In practice: a $0.04 run executed 5,000 times a month is a $200 line item, and the multiplication is where small waste becomes real money.

Beyond inference there is Total Cost of Workflow Ownership: the full cost of owning a workflow, including governance work, infrastructure, and rework when runs fail. In practice: two workflows with the same token bill can have very different total costs once failure rates differ.

Ways the bill shrinks or hides

Prompt caching lets a provider recognize that the start of your prompt is identical to a recent call and charge a reduced rate for the repeated tokens. In practice: a workflow that resends the same 4,000-token system prompt on every run can cut its input cost sharply.

The Agentic cost stack is the set of meters a vendor may bill for agent work: the inference meter (tokens), the application meter (per action, per conversation, or per work unit), and the governance meter (platform or compliance charges). In practice: two platforms doing the same task can bill completely different line items.

Some platforms use Credit-pool pricing: a monthly pool of abstract credits with no published token conversion. Without the conversion, the true unit cost of a task cannot be computed from the outside.

The Cromus Lens

The Cost Simulator computes per-run and monthly cost from verified registry prices across five cost-quality modes, detects prompt-caching opportunities, and models vendor meters beyond inference, such as Salesforce agentic work units, Copilot credits, and ServiceNow governance charges. When a platform is opaque, for example a credit pool with no token disclosure, Cromus reports "unavailable" instead of fabricating a number. A cost model you can trust must be allowed to say it does not know.

Try it below: run the same scenario in two cost modes and watch the per-run number move.