Skip to main content
Control AI agent costs with separate license and usage budgets, enforced execution limits, and total cost per accepted outcome. Athena adds per-run dollar and model-call limits, shared across sub-agents.

​
What are we paying for: seats, tokens, or agent actions?

Seats buy access. Tokens price model input and output. Credits price agent activity. Forecast each separately, then combine them in the workflow budget. These US list prices provide a starting point: OpenAI bills ChatGPT subscriptions and API usage separately. Google also has separate, model-specific Gemini API usage pricing. For each model call, multiply uncached input, cached input, and output tokens by OpenAI’s respective token rates. Divide the sum by one million. OpenAI’s GPT-5.4 Mini charges 0.75permillioninputtokens,0.75 per million input tokens, 0.075 for cached input, and 4.50foroutput.Asyntheticcallwith20,000uncachedinputtokensand2,000outputtokenscosts4.50 for output. A synthetic call with 20,000 uncached input tokens and 2,000 output tokens costs 0.024 at these rates. One thousand identical calls cost 24.Onethousandtaskswithtencallseachcost24. One thousand tasks with ten calls each cost 240. That difference makes call counts essential to forecasting.

​
What makes an agent loop expensive?

Imagine an agent retrying a failed lookup, rereading a file, then asking another agent to investigate. The second agent repeats the lookup. Neither produces an accepted result, but billable attempts continue. OWASP treats uncontrolled model usage as an unbounded-consumption risk. It recommends quotas, timeouts, monitoring, and limits on inputs and actions. Give each task an allowance below the monthly cap. Parallel workers should draw from one shared budget. Test concurrent calls before running a large batch. Record retries against the original task. Otherwise, splitting failed work into new runs can hide its total cost.

​
How do we set an AI budget that stops spending?

  1. Name the work and owner. Define one accepted output, expected monthly volume, and the person who approves spending.
  2. Set several limits. Cap dollars per task and batch. Also limit calls, retries, elapsed time, and concurrent workers.
  3. Test the stop. Run a deliberately failing task with a small allowance. Check whether execution stops, pauses for approval, or only sends an alert.
  4. Test shared work. Repeat with parallel workers and unavailable billing information. Check the total charged to the task.
  5. Bound recovery. Stop attempts that repeat an error without new evidence. Require the owner to approve further work.
  6. Reconcile weekly. Record owner, model, usage cost, retries, review time, and acceptance. Match totals to invoices and investigate expensive failures.
OpenAI’s Agents SDK stops when execution exceeds its max_turns limit. Use it to bound iterations, alongside a separate dollar allowance. Choose one recurring workflow for the first trial. Use the same inputs and acceptance criteria when comparing cheaper models.

​
How do we prove AI ROI without inflating savings?

Measure the manual and AI-assisted workflow against identical acceptance criteria. Include routine items, difficult items, and failures. The workflow owner should review quality. Track staff time separately from elapsed time. A faster result can still require substantial review. Use three measures:
  • Cost per accepted outcome: Total workflow cost divided by accepted outputs. Include failed runs in the cost.
  • Net benefit: Manual baseline cost plus separately evidenced additional benefit, minus total AI workflow cost.
  • ROI: Net benefit divided by total AI workflow cost, multiplied by 100. State the period and whether benefits represent capacity or cash.
Consider a synthetic monthly example with 200 reports. Manual preparation takes 45 staff minutes each. AI-assisted preparation and review take 15 minutes each at unchanged quality. At an assumed 60hourlylaborcost,manualworkcosts‘150×60 hourly labor cost, manual work costs `150 × 60 = 9,000‘.AI−assistedlaborcosts‘50×9,000`. AI-assisted labor costs `50 × 60 = $3,000`. Add 1,500forlicenses,usage,support,andallocatedsetup.TotalAIworkflowcostis1,500 for licenses, usage, support, and allocated setup. Total AI workflow cost is 4,500. Assume no other baseline costs or revenue benefit. If all 200 reports are accepted:
  • Cost per report is $4,500 ÷ 200 = $22.50, versus $45 manually.
  • Capacity-valued cost reduction is $9,000 − $4,500 = $4,500.
  • Capacity-based ROI is $4,500 ÷ $4,500 × 100 = 100%.
The workflow releases 100 staff hours. Record how the team uses them, and retain baseline samples, reviewed outputs, time records, and acceptance decisions. Remeasure after changing the model, workflow, or workload. A successful trial needs a measurement period and an owner for the assumptions.
A failed run keeps adding cost without a run cap; Athena pauses before attempt 6 exceeds its model spending cap and shares one budget across workers. In an example month, 200 accepted reports at the same quality cost $9,000 manually or $4,500 with AI, including labor and other costs, giving $22.50 per report, 100% capacity-based ROI and 100 freed hours rather than cash savings.

​
How do chat assistants, copilots, and Athena compare?

Compare the allowance, stopping behavior, reporting, and model choices for your workflow. Microsoft’s Copilot Studio handles credit-based agent activity; Google’s Gemini in Workspace provides AI within office apps.
AI cost controls, compared by question. Examples: Chat assistant, ChatGPT Business; In-app copilot, Gemini in Google Workspace; Athena. What limits spending? Chat assistant: Workspace spending controls; In-app copilot: Feature and model limits; Athena: Limits on dollars or model calls. What does the budget cover? Chat assistant: One workspace; In-app copilot: One user’s feature use; Athena: One run, shared across sub-agents. Which models can we use? Chat assistant: Models included with your subscription; In-app copilot: Models included with your edition; Athena: Fixed model or selection for each task. Can we see usage? Chat assistant: Usage reporting; In-app copilot: Usage reporting; Athena: Usage reporting.
OpenAI provides ChatGPT Business usage reporting, Microsoft provides Copilot Studio usage reporting. Google provides Gemini usage reporting, and Athena provides agent usage reporting. Athena combines shared per-run limits with cost views by workflow, agent, user, and project. Set the allowance when you start an Agent Operating Procedure run through the API. Add run_budget to the /api/v0/aop/execute or /api/v0/aop/execute-async request with max_cost_usd, max_model_calls, or both. Before each model call, Athena reserves expected cost against the shared allowance. Parallel agents cannot reserve the same remaining budget twice. At the cap, Athena stops the run before the next model call and marks it as not succeeded. Work completed so far is saved. To continue, start a new run with a larger run_budget. Athena can pin a model or route repeated work to the cheapest model meeting the required output-quality standard. Outputs record which model ran. Use Athena’s budgets and governance controls and model pinning and routing to plan the trial. Join usage records with staff time and acceptance decisions for the ROI calculation.

​
What to plan for

  • Unfinished work: Athena’s cap can leave a job incomplete. Assign an owner to accept partial output or approve more budget.
  • Different stopping rules: OpenAI’s alerts leave traffic running; monthly hard API limits can allow slight overages during enforcement delay.
  • Credit thresholds: Microsoft’s Copilot Studio prepaid custom agents stop at 125% capacity; new agent-flow runs stop at exhaustion. Ongoing work can finish.
  • Included usage: Microsoft includes qualifying employee-facing Studio usage for authenticated Copilot license holders, subject to conditions and fair usage.
  • Google limits: Google’s Workspace AI access can be reduced at limits. Google’s Gemini API account and experimental project caps can allow delayed-enforcement overages on spending caps.
  • Budget coverage: Athena’s dollar cap covers model spending; call limits bound calls. Execution stops if its spend tracker is unavailable.
  • Additional charges: Include tools, storage, and OpenAI’s GPT-5.4 Mini regional-processing uplift of 10% where applicable.
  • Savings: Count review, rework, setup, and external services. Released hours become cash savings only when spending falls; avoid double counting benefits.

​
Questions teams ask

OpenAI offers custom Enterprise pricing through sales, with credit-based and token-based arrangements. Obtain a quote covering included usage and billing terms.
Microsoft’s enterprise offer lists $30 per user monthly, paid yearly, plus a qualifying Microsoft 365 license.
Price every call’s input, cached input, and output tokens at OpenAI’s model-specific token rates. Include retries in your forecast.
OpenAI’s ChatGPT Business offers workspace spend controls, while Microsoft’s Copilot Studio uses credit capacity and spending rules. Test the stopping behavior with a small failing task.
Athena shares one per-run dollar and model-call allowance across sub-agents.
Start with total cost per accepted outcome against the manual baseline. Report quality, review time, and failure rate alongside it.
See it work