How to manage AI investments in the agentic era

OpenAI’s goal is to make AI more accessible, capable and affordable over time. From GPT‑4 to GPT‑5.4, the price per million tokens fell 97%. GPT‑5.6 ⁠ continues that progress, delivering better performance in the Artificial Analysis Coding Agent Index with 54% fewer output tokens and 57% less time per task.

But token price alone does not show whether AI is creating value. Leaders should look at useful work per dollar : tasks completed, time saved, decisions improved, and workflows ready to scale.

As teams move from chat to longer-running workflows, admins need clearer visibility into demand, spend, and risk.

For priority deployments, OpenAI’s AI Deployment Engineers ⁠ (opens in a new window) can work directly with customers on evals, architecture, latency, reliability, and workflow design to improve both performance and cost efficiency. Privacy and governance should be part of that work from the start: sensitive workflows need the right access controls, retention posture, compliance visibility, and approval paths before they scale. Where applicable, OpenAI’s enterprise privacy controls, including Zero Data Retention ⁠ (opens in a new window) options, can help customers deploy AI in high-trust environments.

Enterprise leaders need a plain view of AI usage: who is using it, which products or models they are using, how much capacity they are consuming, and what kind of work that usage supports. Without that visibility, a growing bill is hard to interpret. It could reflect waste, productive experimentation, or a workflow that is starting to become business-critical.

ChatGPT Work ⁠ supports longer, multi-step tasks, so usage can vary widely by workflow. Admins need to see the work behind that usage, not just the credits consumed. This is possible thanks to a shared view of demand across ChatGPT. Updated usage analytics and spend controls ⁠ in the Admin Console ⁠ (opens in a new window) help admins see adoption, credit usage, and spend by user, product, and model; track trends over time; identify emerging patterns; and understand when usage reflects broad adoption, a power-user workflow, or a recurring business process that may deserve more investment.

Model choice is only part of the equation. Clear instructions, focused tools, reusable context, and explicit stopping conditions can reduce loops and wasted spend. The goal is to match the model and workflow to the task: use smaller or faster models when they meet the quality bar, and reserve frontier intelligence for complex, ambiguous, or high-stakes work.

Team and user: Where is demand growing, and who may need more support?

For priority workflows, track cost per accepted outcome . In customer support, that might be a resolved case. In engineering, it might be a tested change that passes review. Pair that cost with business value such as time saved, cycle time reduced, revenue protected, risk avoided, or capacity created.

This becomes more important as teams adopt plugins, connectors, Computer Use, and other frontier capabilities that can operate across enterprise systems. ChatGPT Work ⁠ gives admins centralized controls for access, approved context, connected tools, permitted actions, usage, and spend. Spend controls such as workspace defaults, group limits, individual overrides, and review requests with project context help leaders support high-value work without raising limits broadly.

Product and model: Where is more expensive intelligence being used, and is that demand sustained?

The lowest token price does not always produce the lowest total cost. A cheaper model may fail, retry, or create work that needs correction. A more capable model may cost more per token but reach an acceptable result faster, with fewer attempts and less review.

Evaluate models on the work they need to perform. Use evals that reflect real tasks ⁠ , including edge cases, and define “good enough” before testing. Then measure the full cost of reaching that standard: model and tool usage, attempts, completion rate, latency, and human review.