AI COST GOVERNANCE
When token budgets become enterprise risk
OpenAI API pricing is transparent. Enterprise consumption rarely is.
One of the most urgent concerns I am hearing from organizations adopting generative AI is that departments can consume their entire monthly AI budget in a matter of days. The initial reaction is often surprise: individual prompts appear inexpensive, usage does not feel excessive, and no single team believes it is responsible for the increase. Yet the invoice shows that consumption has accelerated far beyond the original plan.
This is rarely caused by one careless employee or one unusually expensive request. It is usually the cumulative result of model choice, large context windows, long outputs, repeated iterations, agent calls, retries, growing user adoption, and applications that quietly send far more tokens than employees realize. If no one owns these multipliers, a low unit price can become an unexpectedly large operating expense.
The current price spread changes the control question
As of August 6, 2026, OpenAI lists standard short-context rates per one million tokens ranging from $0.20 input and $1.20 output for GPT-5.6 Luna to $30 input and $180 output for GPT-5.5 Pro and GPT-5.4 Pro. GPT-5.6 Terra is $2 input and $12 output; GPT-5.6 Sol is $5 input and $30 output. Cached input is priced lower, while cache writes, long context, regional processing, and faster service tiers can change the economics.
The operational lesson is not simply to choose the cheapest model. It is to route each workload to the least expensive model and service level that meets an explicit quality, latency, privacy, and reliability requirement.
How a budget explodes
Consider an illustrative deployment at a midsized firm with 200 employees. Each employee completes 10 AI-assisted tasks per hour—80 tasks during an eight-hour workday. Each task triggers eight model calls, and each call processes 8,000 input tokens and generates 2,000 output tokens across 22 working days. The company may think it is supporting only 200 users. In practice, that workload produces 22.528 billion input tokens and 5.632 billion output tokens in a single month.
- GPT-5.6 Luna: approximately $11,264 per month, or $135,168 per year.
- GPT-5.6 Terra: approximately $112,640 per month, or $1,351,680 per year.
- GPT-5.6 Sol: approximately $281,600 per month, or $3,379,200 per year.
These figures apply the published standard short-context rates and assume a consistent workload for 12 months. They exclude tool charges, storage, regional uplifts, cache writes, networking, and application infrastructure. They are workload illustrations, not forecasts.
Ten AI-assisted interactions in an hour can be a conservative assumption when iteration is frictionless. Consider a law-firm clerk using AI to summarize a court opinion. The first response may overlook an important issue, so the clerk asks the model to emphasize that point, restore omitted facts, locate supporting passages, add citations, reorganize the analysis, compare the holding with another case, and refine the final format. Ten iterations can occur within a single hour on one assignment. The clerk sees a faster path to a better work product, but may not see that each iteration can cause the application to process the source document, instructions, conversation history, and previous outputs again. When the cost of each interaction is invisible to the employee, repeated consumption feels free.
The largest surprise is often not employee adoption. It is amplification inside the system: agent loops, repeated retrieval, full conversation history, verbose outputs, automatic retries, duplicate workflows, and test traffic that reaches production credentials. Each layer can multiply the next. Midsized companies feel this pressure quickly because an unplanned six-figure AI expense competes directly with hiring, product development, and other operating priorities. Smaller companies and startups can feel the pinch even sooner: their absolute usage may be lower, but limited budgets leave much less room for uncontrolled consumption, architectural mistakes, or an unexpected invoice.
Govern the workload, not just the invoice
A monthly vendor limit is a final backstop, not a cost-control strategy. By the time it triggers, the organization knows money was spent but may not know which business outcome, team, workflow, or design decision caused it. Effective AI FinOps connects every request to an accountable workload.
- Attribute: tag usage by business unit, product, environment, workflow, model, and owner.
- Budget: set monthly dollars and per-task token envelopes before production approval.
- Enforce: cap output, context, agent steps, retries, concurrency, and permitted models.
- Route: use smaller models by default and escalate only when measured quality requires it.
- Observe: alert on burn rate, cost per successful outcome, cache performance, and abnormal loops.
- Review: require a named owner and expiration date for exceptions.
Make unit economics a release requirement
Every production AI workflow should have a baseline cost per successful business outcome, not merely cost per request. A support agent might be measured per resolved case; a document workflow per approved file; a research assistant per accepted brief. This makes optimization honest: reducing tokens is valuable only if quality, safety, and completion remain acceptable.
OpenAI pricing will change, and workloads will evolve faster than annual budgeting cycles. Durable governance therefore stores prices as configuration, measures actual token paths, forecasts from observed demand, and tests cost controls alongside quality and security controls. The goal is not to suppress useful AI. It is to make consumption visible, intentional, and economically accountable.
Pricing checked August 6, 2026. Source: OpenAI API pricing. Verify current rates and service-specific charges before making purchasing or budgeting decisions.