What Matters
An AI agent's token bill is only one layer of its cost. Compare model, runtime, tools, memory, payments, verification and human intervention across every attempt, then divide by outcomes accepted inside the same policy. DN's calculator makes that arithmetic explicit. It models your assumptions; it does not establish vendor superiority or production reliability.
DN Evidence Block
- Source review: 19 October 2026. AWS's 8 October payment integration account, IBM's 15 October decision architecture announcement and LangGraph's persistence documentation.
- DN contribution: a cost ledger, an accepted-outcome calculation and a proposed comparison method.
- Evidence type: documentation analysis and hypothetical arithmetic. No paid deployment, live trading, head-to-head runtime benchmark or customer ROI result is reported.
- Author: Heath Muchena. Corrections: contact DN.
The cheap agent can be the expensive workflow
A model generates an answer for a few cents. A paid data call adds another charge. A failed tool request triggers a retry. A person spends ten minutes finding out whether the external action finished. The bill is spread across providers, but the business purchased one outcome.
The wrong denominator hides this. Dividing by model calls rewards activity. Dividing only the costs of successful runs discards the failed attempts that consumed budget. Counting an unverified completion message as success can make a broken workflow look efficient.
Define an accepted outcome before testing. An invoice reconciliation might require a matching authoritative balance, no unauthorized disclosure and an explanation of unresolved entries. A transaction-preparation task might require a valid unsigned payload inside the user's limits. An executed payment needs its own settlement and authorization checks. These are different products, even if the same model performs them.
The DN Agent Cost Stack
| Layer | Include | Accounting trap |
|---|---|---|
| Models and decisions | All billed model work: input, output, cached tokens where charged, routing and specialist decisions | Counting an adapter call as free or omitting failed-run inference |
| Runtime and compute | Hosted execution, sandboxes, local hardware allocation, energy and idle capacity | Comparing a marginal API bill with a fully allocated local system |
| Tools and paid data | Search, external APIs, retrieval services and paid resources | Assuming one resource charge per task when retries repeat it |
| Memory and storage | Persistent state, vector storage, checkpoints and relevant transfer charges | Ignoring retention or billing the same storage in runtime and memory |
| Payments | Applicable network, settlement, conversion or payment-service fees | Adding the purchased resource price here after already counting it under tools |
| Verification and operations | Automated grading, monitoring, security services and incident operations attributable to the workload | Calling an output correct without pricing the check |
| Human intervention | Approvals, review, exception handling, reconciliation and remediation time | Counting only hands-on prompting |
| Allocated setup | Implementation and migration cost spread over an explicitly chosen planning horizon | Presenting an allocation assumption as a cash payment this month |
DN calculation: period cost = the six spend categories above + human hours × hourly cost + setup cost ÷ allocation months. Cost per accepted outcome = period cost ÷ accepted outcomes. All costs across all attempts belong in the numerator. With zero accepted outcomes the ratio is undefined. Report the total spend and failed or unresolved workload instead.
DN Cost Per Accepted Outcome Calculator
Use totals for the same month and workload. Defaults are hypothetical, not market prices. Enter zero only when a category is genuinely zero; missing or negative inputs are rejected. Retries belong in the spend totals. Use one currency throughout; changing the currency label does not convert amounts.
Illustrative total: USD 130.00; USD 1.63 per accepted outcome, rounded. See the worked example below.
The calculator runs locally without credentials, wallet access, model calls or storage of your entries. Normal DN website analytics are separate. The CSV contains your entered scenario, so share it only if you intend to share those numbers.
A worked example that includes human rescue
Hypothetical scenario: 100 attempts produce 80 accepted outcomes. Models, runtime, tools, storage, payment fees and automated assurance total USD 50. Review and rescue take 1.5 hours at USD 40 per hour: USD 60. Setup of USD 120 allocated across six months adds USD 20 for this monthly scenario. Total: USD 130; cost per accepted outcome: USD 1.625 before rounding.
The direct technical spend alone would suggest USD 0.625 per accepted outcome. The difference is the work that made the result usable. This example establishes arithmetic, not that these inputs represent any provider. Human time is often estimated; label estimates, record their basis and replace them with observations when available.
Changing the allocation horizon changes the displayed unit economics without changing the original setup invoice. Keep cash outlay, ongoing run cost and planning allocations visible separately. Do not add total setup cost and its allocation to the same period.
Why the architecture changes the bill
AWS's 8 October 2026 account describes BlockRun and Incarna using x402 for paid model inference with wallet connectivity and infrastructure spending controls. This is a vendor-described deployment example, not a DN cost benchmark. For accounting, separate resource consumption from the cost of paying for it.
IBM's 15 October Granite Switch announcement describes decision adapters alongside chat and speech capabilities with shared cache behavior. Specialized routing or decisions can change the cost structure, but the full workflow must establish whether any saving survives verification and failures. DN has not reproduced IBM's performance claims.
LangGraph's persistence documentation distinguishes thread checkpoints from cross-thread stores and discusses checkpoint retention. Saved state has operational and storage implications. Merely having a checkpoint does not establish the correct external outcome after resumption.
Managed, local or hybrid: compare the actual deployment
| Configuration | Best fit to investigate | Costs and access to verify | Avoid assuming |
|---|---|---|---|
| Managed | A team that needs hosted execution and integrated operations | Runtime billing, tool fees, quotas, region, data handling and export access | That platform bundling includes every external tool or eliminates human review |
| Local/self-hosted | A team with suitable hardware and deployment skills | Hardware allocation, utilization, energy, maintenance and model-license terms | That a downloaded model means zero cost or that connected tools keep data local |
| Hybrid | Workflows that can route bounded tasks to specialists and escalate exceptions | Routing overhead, cross-system traces, duplicated state and policy consistency | That switching models preserves permissions, context or acceptance quality |
These configurations describe execution, not custody. A local process can hold powerful remote credentials; a managed process can be read-only. Inspect signing authority, tool permissions and revocation separately. Eligibility, regional access and operational status must be checked for the exact service and account before procurement.
Turn the calculation into a defensible experiment
- Freeze the workload, authority boundaries, acceptance rules and accounting period.
- Pin model, decision component, framework, tool and configuration versions.
- Record every attempt and its completed, rejected or unresolved outcome.
- Use authoritative state checks and an independent review of the grader.
- Keep attributable costs in exclusive categories; disclose shared-cost allocations.
- Repeat comparable trials. Publish sample size, variability, omissions and the cost of exceptions.
Report policy violations separately. They are not a surcharge that can be traded against cheap success. A configuration outside the required authority rules is ineligible for that workflow regardless of its average cost.
Put the evidence to work
Use the calculator to find the cost category worth investigating, then use DN's Model or Harness Investigation Planner to choose the next experiment. Inspect Agent Trust evidence and the Managed Runtime Index before comparing implementation options. For a defined research question, contact DN about a scoped comparison. This is a research enquiry, not an assurance certification.
DN Alpha Thesis
The useful economic unit is a verified, policy-compliant result. An operator that can trace technical spend, human intervention and external outcomes has a stronger basis for buying automation than one optimizing a token-price table. This is an editorial thesis to test. It does not predict that one runtime, model family or payment network will capture the market.
Methodology, limitations and change log
Version 1.0, prepared 19 October 2026. The calculator sums user-entered period totals and an explicit setup allocation. It does not fetch live pricing, estimate success probabilities, convert currencies, calculate tax or model incident-loss distributions. It does not distinguish cash and non-cash costs unless the user allocates them correctly. Missing costs make the result incomplete. Unlimited precision is not assurance: outputs are rounded for display and remain scenario estimates.
No original provider performance dataset is claimed. The original contribution is the accounting method, worked example and local tool. Validation covers arithmetic and edge cases, not the accuracy of a user's inputs. Corrections should identify the input assumptions, expected result and relevant source.
Primary sources
- AWS: pay-per-inference with BlockRun and Incarna, 8 October 2026
- IBM: Granite Switch 4.1, 15 October 2026
- LangGraph: persistence documentation
Frequently asked questions
Is cost per accepted outcome the same as token cost?
No. It includes attributable operating and human costs across all attempts and divides by outcomes accepted under defined rules.
What happens when no outcome is accepted?
The ratio is undefined. Report spend and unresolved or failed attempts instead of displaying zero cost per success.
Does choosing ZAR convert the example?
No. The currency selector changes the label only. Enter all costs in the same currency using your own accounting basis.
Does a lower cost mean a system is safe?
No. Required authority, security and acceptance conditions are separate eligibility requirements.
Can I compare a managed system with a local model?
Yes, if the task and accounting basis match and you disclose hardware, human time, shared allocations and unavoidable configuration differences.
Disclosure: educational research and scenario modelling, not personalized financial, accounting or procurement advice. No affiliate signup is required to use this tool. Links in this edition are research and DN navigation links; no referral code or partner ranking is embedded.
