OpenAI Starts Charging Only When the AI Actually Works

For a decade, AI pricing has been a meter running in the background: seats per month, tokens per million, compute per hour. On August 31, The Information reported that OpenAI has spent recent months quietly offering some of its major customers a different deal — one where they pay only when the AI actually completes a task. It is a small experiment with a large implication: the largest AI vendor on the planet is testing whether the industry’s unit of value should stop being usage and start being results.

What was reported

According to The Information’s briefing, OpenAI has begun giving select major enterprise customers the option of paying on outcome — the AI does the work, the work gets verified as done, and only then does money change hands. The report frames it as part of an “industrywide movement toward outcome-based pricing,” one that could reshape how AI contracts are structured as agents take on jobs previously measured in human hours.

OpenAI has not published pricing tiers for the model, and the arrangement appears to be negotiated per customer rather than listed on a rate card. That in itself is telling: outcome pricing is inherently bespoke, because no two customers define “a completed task” the same way.

The shift from usage to outcomes

To see why this matters, it helps to look at the pricing ladder the industry has climbed. First came per-seat pricing — the SaaS default, imported unchanged into AI products. Then per-token pricing, which at least tracks real consumption: Zylo’s data puts the average organization’s OpenAI API spend at roughly $384,500 a year as of April 2026, a number driven almost entirely by token meters. Then hybrid models — usage plus seats plus credits — which Salesforce now runs in at least three parallel variants for Agentforce.

Outcome-based pricing breaks the ladder entirely. The customer pays a fixed amount per resolution, per completed procedure, or per verified result. If the agent fails, the vendor eats the compute cost. Vendors like Sierra have been explicit about the philosophy: “With outcome-based pricing, Sierra gets paid only when we complete a task for you.”

The template was set in customer support, where outcomes are easiest to define. Intercom’s Fin agent launched at $0.99 per resolution in 2023 — charged only when a customer conversation is closed end-to-end — and the model has since spread to Zendesk, Sierra, Legora, AirHelp and others. Sierra’s outcome rates start at $1.50 per automated resolution, or $2 on pay-as-you-go. Salesforce’s reported ~$3.6 billion acquisition of Fin is, among other things, a bet that per-resolution economics scale: when Salesforce CEO Marc Benioff says Agentforce customers can now negotiate contracts tied to revenue growth or cost savings, he is describing outcome pricing negotiated at the enterprise layer.

Agentforce itself has been the highest-profile test bed. SaaStr’s analysis of Salesforce’s three-plus Agentforce pricing models — consumption credits, flex credits, and now outcome-tied custom contracts — reads like a live A/B test of the entire pricing stack, with outcome terms increasingly winning the enterprise conversations.

Why OpenAI is doing this now

Three forces converge on OpenAI’s experiment.

First, agents changed the product. A chatbot bills naturally by token; an agent that plans, calls tools, and completes a multi-hour workflow does not. When the deliverable is “the refund was processed” or “the report was filed,” customers stop thinking in tokens and start comparing against the cost of a human doing the same job — often $20 to $60 per task. Per-token pricing becomes almost incomparable at that point.

Second, the promotional-pricing era is ending. This very week, the industry is repricing itself: Claude Sonnet 5’s $2/$10 introductory rate expired on August 31 and reverted to $3/$15, GPT-5.6 Sol’s promotional $4/$20 is scheduled to revert to $5/$30 around November 21, and Claude Code’s summer usage boost winds down September 14. The land-grab phase — capability priced to acquire users — is closing. What replaces it has to justify real margins, and “we only charge when it works” is a much easier margin story to sell a CFO than another token-price increase.

Third, the idea has influential sponsors inside OpenAI’s orbit. Bret Taylor — OpenAI board member and Sierra co-founder — has spent 2026 arguing publicly that outcome-based pricing is the future of software business models and that the “atomic unit” of AI value is the completed task. When the board’s own framework gets piloted with major customers, the direction of travel is hard to miss. Andreessen Horowitz has been pushing the same thesis since late 2024, arguing that agents hand startups a pricing weapon incumbents can’t easily copy because legacy vendors’ revenue depends on seats.

The hard part: attribution

The most sobering caution comes from the payments layer. Stripe — which sees more AI-vendor billing than almost anyone — has warned that attribution in outcome pricing will be messy, because business outcomes rarely trace cleanly to one AI system. A support ticket resolved by an agent may have been escalated by a human, unblocked by a knowledge-base article another vendor wrote, and closed only because a third-party logistics API finally returned the right answer. Whose outcome was it?

That question multiplies in multi-agent systems, where several vendors’ models collaborate on a single result. It also creates the classic measurement traps: Who verifies that a task was “completed”? Can vendors game the definition? Do customers dispute outcomes the way they dispute invoices today — and who arbitrates? Outcome pricing moves the battleground from unit economics to contract language, and the companies that win may be the ones with the best auditors, not the best models.

What it means

For buyers, the shift is initially good news: pilot risk drops to near zero, since failed agents cost nothing, and budget conversations move from “how many tokens did we burn” to “what did it cost per outcome versus the human baseline.” Expect procurement teams to start demanding outcome terms in every AI contract simply because they can.

For AI vendors, it is a margin squeeze with a filter effect. Charging per outcome means internalizing the cost of every failed attempt — every hallucinated step, every retry, every task a human has to finish. Vendors confident in their agents win; vendors shipping 70% reliability discover that outcome pricing converts their error rate directly into a revenue haircut. The model’s center of gravity will likely stay in domains where outcomes are crisp and verifiable — support resolutions, document processing, coding tasks with passing tests — before anyone dares price fuzzier knowledge work this way.

And for the broader software industry, OpenAI’s experiment is the mainstreaming signal. When per-resolution pricing was an Intercom novelty, incumbents could ignore it. When it is a negotiation lever available from OpenAI itself — the vendor whose API sits underneath thousands of products — every SaaS pricing page that says “per user per month” is suddenly negotiating against a rival metric. Salesforce’s Benioff offering revenue-tied contracts and OpenAI testing pay-per-task within the same month suggests 2027’s software budgets will be written in a different currency than 2025’s.

The token meter isn’t going away — most API traffic will bill on usage for years. But the industry’s most important price is no longer “$X per million tokens.” It’s “what do you pay when the machine actually does the job?” OpenAI just started answering that question with real contracts.

---

Sources are listed in the article metadata. Reported figures verified against The Information’s August 31 briefing, Sierra and Intercom’s published pricing, Stripe’s outcome-pricing guidance, and SaaStr’s Agentforce pricing analysis.