The Price of Intelligence - AI Agent Pricing in 2025

Should an AI agent charge for access, activity, or a completed job? The answer depends on what it does and who buys it. Kshitij Grover introduces this problem from his perspective as co-founder and CTO of Orb, a usage-based billing infrastructure company serving AI, infrastructure, and developer-tool businesses. Customer support, sales, enterprise search, and enterprise agents have different purchasing patterns; a single pricing formula will not fit them all.

Intercom’s Fin makes the choice concrete. In the pricing page shown, Essential, Advanced, and Expert plans provide a familiar tiered structure. The Fin layer adds $0.99 per resolution. That connects the incremental charge to the support work the agent completes, and Grover reads it as a signal of confidence in the product. The billable definition of success matters: Intercom’s historical explanation of Fin includes both affirmative customer confirmation and a customer leaving without requesting a human. A billable resolution therefore does not always mean an explicitly confirmed answer. The pricing pages throughout this article are the talk’s historical examples, not current purchase terms.

Unify’s go-to-market product combines Growth, Pro, and Enterprise tiers with a pricing calculator. Beneath the tiers sit credits, usage limits, and caps. Seats can be included in a tier while individual features carry usage-based or hybrid charges. The customer is choosing a package, but the billing system must account for several independent dimensions.

Cursor’s shown plans are free, $20 per month, and $40 per month. Those headline prices conceal distinctions between completions and requests, fast and slow service, and premium models. GPT-4, GPT-4o, and Claude 3.5 Sonnet appear in that premium category, with different limits or costs. Even a tool aimed at everyday developers can require customers to understand more than a subscription price.

Chargeflow takes a more direct route to value: it charges a percentage of a successfully recovered chargeback. Payment is contingent on recovery, so Grover describes the arrangement as an effective ROI guarantee. The useful distinction is that the fee follows recovered value; it does not guarantee that every dispute will be recovered or that the customer’s overall business will be profitable.

These examples apply established pricing principles to a new product category. Start with the buyer and the process around them. An individual developer at a small business might enter a credit card and start working. A Fortune 100 procurement team may need to budget the agent against existing, more traditional software. The same pricing page can help one buyer and obstruct the other.

Simplicity and predictability should rarely be compromised. Usage pricing makes predictable spending particularly important: understanding today’s unit rate is not enough if the customer cannot estimate next month’s bill. Pricing also shapes expansion after the initial purchase. A meter can encourage workloads the product handles well—or discourage the very usage that would make it valuable.

Clay’s pricing page communicates its intended audience through both packaging and presentation. Grover points to the emphasized Explorer and Pro tiers, then to people/company search quantities of 10,000 and 25,000. Those quantities suggest a substantial prospecting workflow. The OpenAI, Airbnb, Anthropic, and Canva logos reinforce a story about fast-growing companies using the product, rather than a page aimed only at traditional enterprise procurement.

Replit tells a different story. Its mission of making programming broadly accessible is reflected in a lower entry price and a prominent free tier. Yet the page is technically specific: agent checkpoints, Autoscale, and deployment add-ons introduce granular billing concepts. Accessibility does not necessarily mean eliminating technical detail; the detail still has to make sense to the intended user.

A page with no visible price and only a demo request can be appropriate too. HEAVY.AI directs enterprise buyers into a sales conversation rather than a self-service tier selection. ServiceNow’s AI-agent page does something similar. From its enterprise logos and existing customer relationships, Grover infers that ServiceNow is cross-selling agents into accounts it already serves. A salesperson belongs in that workflow; an individual developer signing up independently is not the target.

An AI sales development representative introduces a different comparison. The example page pairs a flat monthly rate with emails sent and meetings per month. A buyer can compare those outputs with the productivity of an existing SDR team and ask whether to hire another person or adopt the agent. The pricing presentation helps make that staffing decision because it supplies units the buyer already uses to evaluate the work.

Cognition’s Devin is also presented as a teammate, but its pricing introduces a translation step. The talk’s example combines a monthly charge with agent compute units, or ACUs. These reflect resources such as virtual-machine time, inference, and networking bandwidth. An engineering leader usually thinks about salaries, headcount, milestones, and project costs—not how much networking an engineer consumes.

Grover allows that ACUs may work for Devin. The cost is buyer education: customers must learn how their expected work maps to an unfamiliar consumption unit before they can confidently budget for it.

Grover invokes the Bezos idea that a supplier’s margin creates a competitor’s opportunity. He cites typical SaaS margins of 80–90% as a reference point, then emphasizes that competition can erode them. For agents, today’s input prices are especially unstable. Understand the axes of cost rather than designing the business around a fixed cost snapshot.

Workload variation is a practical starting point. Even within one vertical, some customers use an agent occasionally while others run it intensively. You do not have to preserve the same margin on every interaction, but you do need to identify extreme workloads, decide roughly how to protect against them, and choose which workloads to encourage. Technical R&D can then improve the cost structure instead of leaving margin protection entirely to price increases or tighter limits.

The cost inventory includes models, training, and operations. Operations are easy to overlook, while inference—the recurring work of running the agent model—is usually the major agent operating cost in Grover’s framing. These categories help separate what changes with each use from the investments that can make future use cheaper.

Character.AI connects this infrastructure work to consumer-product viability. Its June 2024 inference optimization post reports approximately 20,000 queries per second, roughly 20% of estimated Google Search request volume. The talk conflates that reported traffic with a separate scenario: 100 million daily users spending an hour each day are assumptions in a future cost calculation, not measured adoption or engagement. Grover dates the company’s early inference optimization work to 2022–2023 and uses it to explain how a service built around sustained conversation can become economically viable.

Jasper presents the same relationship from the customer’s side. Marketers iterating on Brand Voice output do not naturally budget their work in word counts. Unlimited credits on some paid tiers remove that friction and support rapid revision. Grover attributes the economics to a model decision-making engine that selects among OpenAI, Anthropic, and Cohere. That is his causal explanation; the associated announcement establishes unlimited credits and multi-model selection, but does not quantify routing savings sufficient to fund the change.

Pricing cannot be a set-and-forget exercise. Orb calls the ongoing search for alignment pricing product-market fit: as the product improves, the business must keep measuring willingness to pay. Inputs, customer habits, markets, and use cases all change. Additional R&D can increase delivered value, but a rigid pricing model may prevent the company from capturing it.

The talk contrasts abrupt annual or twice-yearly changes with continuous evolution. Large, infrequent jumps can damage customer relationships while leaving new value unpriced between changes. The alternative is to evolve pricing alongside the product and bring customers through those changes, rather than treating pricing as an occasional correction.

Falling inference prices make that flexibility consequential. Grover shows a chart of OpenAI token prices declining over the preceding year to year and a half. Because model usage is an input to many agents, cheaper inference can create expectations of more competitive agent pricing. It can also change which products are viable: he forecasts that data-intensive healthcare and legal applications that look uneconomic in 2024–2025 could become feasible within one or two years. The opportunity is larger than reducing an existing bill; it includes building products whose workloads previously cost too much.

Flexibility extends beyond changing a dollar amount. Luma’s Dream Machine example includes web versus iOS access, relaxed mode, a credit system, and rate limits. Those are separate packaging levers. Adjusting them can serve different customer needs without changing every part of the offer, but each additional dimension consumes some of the buyer’s tolerance for complexity. Maximize useful flexibility within the audience’s need for simplicity.

Prepaid credits are common because they solve several problems at once. First, they bring cash in before consumption generates costs. For a business with substantial cost of goods sold, or COGS, even carrying one month of pay-as-you-go credit risk may be difficult. Requiring payment upfront can also counter fraud by making users commit money before consuming resources.

Credits also simplify discounting across a product with several differently priced SKUs. Instead of negotiating a new dollar rate for every line item, a company can discount the conversion between money and credits while retaining each SKU’s credit charge. The relationship can be expressed as:

The commercial discount changes the second line’s conversion rate; the relative credit prices of the underlying operations can stay unchanged.

An upfront annual allocation also accommodates seasonal demand. Customers can choose when to consume the credits over the year instead of matching each month’s activity to a fixed allowance. Grover sees this mechanism used at both ends of Orb’s customer spectrum: credits support small-company trials as well as commitments in multimillion-dollar deals. The same consumption model can therefore serve very different purchasing commitments.

Looking ahead to 2025, Grover expects two pressures to coexist: price wars in some agent verticals and continuing demands from investors and markets to protect margins. As inputs become more commoditized, he predicts that competition will push companies toward effectively unlimited plans with fewer meaningful usage restrictions. These are forecasts about where the economics could lead, rather than evidence that every agent category can already support unlimited consumption.

He also expects more outcome-based pricing, which makes the contract around success increasingly important. A provider must specify what it guarantees, what its service-level agreement covers, and what counts as a successful result. The unit being sold needs a clear definition before customers can evaluate the offer or audit the bill.

That creates product work as well as pricing work. Grover forecasts more R&D investment in monetization so customers can control and inspect their spending:

• Throttling: Restrict how particular use cases consume the product.
• Spend caps: Set boundaries on expenditure.
• Credit visibility: See how an allotment is being consumed over time.
• Usage auditing: Inspect the activity behind agent charges.

These capabilities make a variable bill understandable while the customer is using the product, rather than only after an invoice arrives.

The implementation burden extends beyond selecting a meter. Enterprise agreements, discounts, and ramps layer additional rules onto pricing logic. Customer visibility has to remain coherent through those variations. Frequent price changes add another difficulty: some customers may still be on earlier price points while others purchase the latest offer.

Grover identifies technical challenges across the full billing path:

Orb is positioned as a billing engine specialized in these demands. The system has to connect what the customer did, what their agreement charges for it, and what finance records—not merely display a price on a landing page.

Grover reports that individual Orb customers sometimes change prices many times a month. That frequency makes price versioning and subscription migrations first-class billing requirements, capabilities also listed in Orb’s billing offering. A pricing change needs a defined version and a way to move customers between versions. Otherwise, the flexibility promised by the product becomes a recurring manual problem in the billing system.