OpenAI asks a direct question: how do you know the AI you buy actually creates value and doesn't just produce a bigger bill?
It's not enough to compare prices per token. What matters is useful work per dollar: cases solved, time saved, better decisions, and workflows ready to scale.
Why price per token isn't everything
From GPT-4 to GPT-5.4, the price per million tokens dropped 97 percent. GPT-5.6 follows that path: in the Artificial Analysis Coding Agent index it improves efficiency with 54 percent fewer output tokens and 57 percent less time per task.
Sounds good, right? But watch out: a cheaper model per token can require more attempts, produce errors that must be fixed, or need human review.
A pricier model per token might solve the task faster and with fewer correction cycles. That's why the right metric isn't just price per token, but cost per accepted result.
Five ways to invest with confidence
- Visibility into who uses what
- Ask yourself: who is using AI in my company, which products or models are they using, and for what kinds of work? Without that visibility, a growing bill could be waste, productive experimentation, or the start of a critical workflow.
- Tools like ChatGPT Work and the admin console provide adoption analytics, credit usage, and spending controls by user, product, and model. Seeing demand at different levels helps decide whether to invest, train, or limit.
- Measure useful work per dollar
- Evaluate models with real tests that include edge cases. Define beforehand what 'sufficient' means for the task.
- Measure the full cost to reach that standard: model and tool usage, retries, completion rate, latency, and human review.
- Example: in customer support measure cost per case resolved; in engineering, cost per change tested and approved.
- Tune model and flow, not just price
- Small changes in prompts, focused tools, reusable context, and explicit stopping conditions reduce loops and unnecessary spend.
- Use smaller or faster models when they meet the quality bar; save frontier intelligence for complex, ambiguous, or high-risk tasks.
- Governance as an operational layer
- Define what context ChatGPT can use, what tools it can access, what actions it can take, and who approves higher-risk steps.
- This becomes more important as you adopt plugins, connectors, and capabilities like Computer Use. Centralized controls let you manage access, approved context, connected tools, and usage limits.
- Spending controls (workspace defaults, group limits, individual overrides, and review requests with project context) help support valuable work without opening the budget to the whole company.
- For priority deployments, implementation teams can help with evals, architecture, latency, reliability, and flow design, including privacy from the start. Options like Zero Data Retention are useful in high-trust environments.
- Treat AI investments like a portfolio
- Distribute budget: broad access for daily productivity, function-specific flows for repeatable work, and strategic bets around your own context.
- Maturation guides funding: exploration to test if a model can do the job; validation with representative cases; production for integrations, controls, and organizational change.
- Invest centrally in shared capabilities: identity, trusted connectors, curated knowledge, evaluations, observability, model routing, and reusable agent patterns.
- Align the commercial structure to usage patterns: Guaranteed Capacity for production systems that need certainty; Scale Tier for predictable high-volume API; Batch API, Flex processing, or Prompt Caching for asynchronous or repeated work. For large deployments, specialized teams can help scale with the right structure.
What you can start doing today
- Request a usage report by workspace, team, and model. If none exists, audit critical workloads manually.
- Define a simple 'accepted result' metric for two high-volume flows (for example, support and an internal process) and measure cost per result.
- Set basic governance rules: who approves tool access, what data the AI may use, and spending limits per group.
Thinking of AI as a portfolio and results per dollar shifts the conversation from 'how much does this cost?' to 'what value does it deliver?'.
That shift is what lets you scale without losing control, and what separates noisy spending from strategic investment.
Original source
https://openai.com/index/managing-ai-investments-in-agentic-era