AI billing does not, at this time, suddenly shift from a per-token pricing model to a pay-per-result model. OpenAI combines three pricing models: tokens for the API, a monthly subscription for ChatGPT, and an economic metric of “cost per successful task.” For an SME project, the real issue is therefore not choosing a trendy model, but understanding what it makes predictable: usage, budget, quality, or accountability.
AI Billing: The Three Models to Distinguish
The per-token pricing model remains the most common way to integrate OpenAI into a website, app, or internal tool. A token is a text segment processed by the model: often a part of a word, sometimes a short word. The more text your user sends, the longer the model’s response, and the higher the bill.
OpenAI announced in 2026 that it would adopt a token-based API pricing model, with, for example, GPT-5.6 Terra priced at $2 per million input tokens and $12 per million output tokens, and GPT-5.6 Luna at $0.20 per input token and $1.20 per output token. Its Scale Tier offering also sells capacity units in tokens, purchased in advance for a given model, with a minimum 30-day commitment.
Second model: the subscription. ChatGPT offers paid plans on a per-user, per-month basis, such as Go, Plus, Business, and Enterprise; the OpenAI Help Center listed ChatGPT Plus at $20 per month in February 2026. This is manageable for a business leader, but it doesn’t mean unlimited access: these plans typically include limits, quotas, or usage budgets.
Third principle: the result. Since 2026, OpenAI has been referring to “cost per successful task,” which is the total cost divided by the number of tasks that meet the expected quality level. At this stage, this is not a public replacement for OpenAI’s token-based pricing. Rather, it is a more meaningful way to measure return on investment.
Why the token doesn't tell the whole story about your budget
The per-token pricing model has one advantage: it aligns with the model’s actual usage. For a support chatbot, a search in a document database, or a summarization feature, this makes perfect sense. You pay for the amount of text analyzed and generated.
Its flaw is less obvious. The token does not measure quality, usefulness, or time saved. A long, incorrect, or unusable response can cost more than a short, relevant one. OpenAI acknowledges this in its 2026 publications: the token price alone does not indicate whether the AI is creating value.
The classic pitfall, from a budget perspective, is estimating the cost based on a ten-question demo. In production, users ask follow-up questions, attach complete documents, request clarifications, and keep the conversation going. In the projects we manage, we often see actual volumes that are two to five times higher than the initial estimates because the usage analysis wasn’t conducted thoroughly.
For a French SME, a simple AI prototype might cost a few hundred euros in tokens per month, but the overall budget lies elsewhere: scoping, integration, testing, security, hosting, and monitoring. Honestly, unless you’re dealing with very high volumes, trying to save a cent per million tokens before you’ve even validated the use case is often a low priority.
Subscription, tokens, results: a practical comparison
The AI billing model changes the way you manage risk. The subscription model puts the accounting department at ease. Tokens put the technical teams at ease. Pay-for-performance puts the business side at ease—but only if the results are unquestionably measurable.
| Model | Real-world examples in 2026 | Benefits for an SME | Main boundary |
|---|---|---|---|
| Price per token | OpenAI API: GPT-5.6 Terra at 2 $ per 1M input tokens and 12 $ per 1M output tokens; Luna at 0.20 $ and 1.20 $ | Pro-rata billing based on actual usage, useful for APIs, chatbots, and automation | Variable budget and no guarantee of quality |
| Subscription | ChatGPT Plus priced at 20 $ per month; Go, Plus, Business, and Enterprise plans | More predictable cost per user | Quotas, usage limits, and limited business integration depending on the plan |
| Reserved Capacity | OpenAI Scale Tier: token units purchased in advance, minimum 30 days | Greater visibility for large API volumes | Commitment and the Risk of Oversizing |
| Result | According to Axios, in 2024, Sierra billed for customer service incidents handled or resolved by its AI agent | Direct alignment with business value | Attribution, Quality Control, and Disputes Over What Counts as “Successful” |
This table hides a simple reality: a single company can use all three models. Subscriptions for collaborators, tokens for a business application, and performance metrics to determine whether the project is worth expanding. Good governance often combines accounting and operational metrics.
When Accrual-Based Accounting Really Works
Performance-based pay is appealing lorwhen a task has a clear endpoint: a resolved customer issue, a qualified sales lead, an invoice matched to an order, or a software test that passes.
In the customer service sector, Sierra was cited by Axios in 2024 as an example of a company that charges for every incident handled or resolved by its AI agent system. The model makes sense: the client company doesn’t want to buy tokens; it wants to reduce the cost of handling requests. But we need to define what constitutes a resolution: an accepted response, no reopening of the case, customer satisfaction, or human validation?
For a sales representative, billing based on leads or appointments seems natural. But it quickly becomes complex. Who assigns the appointment to the AI if a sales rep follows up afterward? What happens if the sales cycle lasts three months? How can you avoid low-quality leads generated just to pad the numbers?
In the context of development, OpenAI cites examples where a task is considered complete only when a code change passes the tests. This is more objective. However, the tests must be sound, there must be a human review, and the costs of rework must be accounted for. An agent that quickly produces fragile code may show a good cost per successful task but create costly technical debt six months later; this point ties in with the trade-offs discussed in our analysis on The Decline in the Cost of Coding Thanks to AI.
Chatbots, Sales Reps, Development: Finding the Right Balance
An information chatbot rarely lends itself to performance-based billing. The success of an open-ended response is subjective: Is the user satisfied? Did they understand? Did they avoid having to call support? In this case, it’s often best to combine a per-token price, a monthly cap, and a satisfaction metric.
An enterprise AI agent operating within business tools requires greater caution. In its 2025 guide on agents, OpenAI recommends providing for human intervention when failure thresholds are exceeded or when the action is sensitive—such as refunds, payments, or order cancellations. A performance-based billing model can lead to poor decision-making if it rewards speed over safety; the issue of confidentiality also warrants clarification, as explained in our guide on Integrating an AI agent without data leaks.
For an AI mobile app, the pricing model depends on the use case. An app that generates short-form content can be billed per token with a cap per user. An app that automates a repetitive work task may be billed based on the time saved, but the vendor’s invoice will often still be based on API usage. Budget estimates are rarely determined solely by the business model: UX, backend, authentication, and maintenance are just as important—and sometimes even more so—as with a AI mobile app for small and medium-sized businesses.
The automated workflow, on the other hand, is the best candidate for results-based measurement. For example: sorting 10,000 incoming requests, extracting relevant informations, and transferring ambiguous cases to a human. Here, we can track the automation rate, the error rate, the time saved, and the cost per processed file.
What You Need to Clarify Before Signing
Before choosing an offer or a service provider, understand the financial mechanics. An AI invoice that isn't properly structured can become unpredictable, but a performance metric that isn't clearly defined can lead to disputes. A good contract outlines both the limitations and the promises.
- Define the metric: resolved conversation, processed document, qualified appointment, closed ticket, validated test.
- Set usage limits: monthly budget, number of users, maximum document length, and number of requests.
- Plan for quality control: human sampling, error threshold, logging (actionable historical data), and correction procedures.
- Clarify the data: GDPR 2018, personal data, retention, hosting, access rights, and data processors.
- Calculate the total cost: tokens, integration, monitoring, labor, rework, supporT, and security.
Security shouldn't be an afterthought. An agent connected to a CRM, Stripe, an ERP system, or an email account has more power than a simple chatbot. At this level, backups, logs, and access rights become safeguards just as important as the choice of OpenAI model; the principles of 3-2-1 Backup Strategy Against Ransomware remain very useful for thinking about resilience.
From the agency’s perspective, the natural approach is to start with a limited pilot: a defined scope, a budget cap, a set of tests, and then a business metric. Generally, four to eight weeks are enough to determine whether AI actually reduces processing time or whether it merely shifts the work to human verification.
Will OpenAI really replace the per-token pricing model?
Nothing in the public information available in 2026 indicates that OpenAI is replacing its token-based AI billing with a generalized performance-based billing model. The ChatGPT pricing pages continue to be organized by user subscription. The API and Scale Tier pages remain structured around tokens and capacity.
The trend is actually going in a different direction: OpenAI is encouraging buyers to focus on the amount of useful work per dollar. This is a healthy development. Two models can have very different prices per token and yet yield the opposite cost per successful task if one performs better, makes fewer errors, and requires fewer prompts.
For a manager, the correct approach is therefore this: continue to request the technical price, but base your decision on business value. A reliable quote should outline volume assumptions, limitations, variable costs, and success metrics. If it promises performance-based billing without explaining allocation, quality, and potential disputes, be wary.
Establishing a clear framework for this type of project early on helps avoid most unpleasant surprises: defining the deliverables, choosing the model, ensuring data security, conducting tests, and setting a budget cap. This is often where an outside perspective can save time—even before development begins.
FAQ on AI Billing
What is token-based AI billing?
This pricing model is based on the volume of text sent to the model and generated by it. It works well for APIs, but it does not guarantee that the response will be useful or cost-effective.
Does OpenAI already charge based on results?
Not publicly as a general model in 2026. OpenAI primarily refers to cost per successful task as an economic metric, while its API offerings remain token-based and its ChatGPT offerings are subscription-based.
Is revenue-based billing better for an SME?
Yes, if the result is objective, verifiable, and attributable to AI. For open-ended responses or long-form sales, it can lead to disputes and requires very specific control rules.
How can you avoid an unpredictable OpenAI bill?
Set usage limits, restrict document length, log requests, and start with a pilot program. Monthly token tracking should be linked to a business metric, not just a cost line item.