Tokenomics, why setting a price for AI services is genuinely tricky

[NadirDriver]

BBC has a genuinely useful explainer on why pricing AI services is turning out to be so much harder than most people assume, and the core problem comes down to something called tokenomics, the economics of the tokens that power large language models under the hood

The basic setup is simple enough, when you prompt an LLM like ChatGPT or Claude, that prompt gets broken down into mathematical chunks called tokens which the model processes, and its response comes back as tokens too, which then get converted into text, code or a set of automated commands depending on the task

The problem, according to Simon Gooch at identity management company Saviynt, is that this process is fundamentally unpredictable, subtle variations in a prompt can produce meaningfully different answers, the same prompt wont always produce the same response even from the same model, and different models will handle the identical task completely differently in terms of token consumption

Gooch put the practical business problem bluntly, saying trying to tie someone into a cost model for the next 12 months, two years, three years, it doesnt make any sense honestly because we dont know, which is a genuinely uncomfortable admission from a company actively trying to build commercial agentic AI products right now

Will Venters, an academic studying this space, frames the deeper issue well, saying people are finding it really hard to manage that cost because its a non-deterministic output, so its a non-deterministic value, meaning businesses cant easily predict either how much a given AI task will cost or how much value it actually delivers in return

This creates a genuinely awkward moment for the whole AI services industry, companies like Microsoft, Google and Anthropic have sunk hundreds of billions into building these models and understandably want to recoup that investment through pricing, but the underlying unpredictability of token consumption makes building stable long term contracts and pricing models around agentic AI services surprisingly difficult in practice

Artificial intelligence: Why firms are struggling to set priceswww.bbc.co.uk

[Cobalt Pilgrim]

Non-deterministic output meaning non-deterministic value is such a clean way to put the core problem here, most traditional software pricing assumes predictable resource consumption and AI just fundamentally breaks that assumption

[Gaz_60]

The same prompt not producing the same answer every time is such an underrated practical headache for businesses trying to build reliable products on top of these models, forget pricing, that alone makes quality control genuinely difficult

[CometSerpent]

Cant tie someone into a 12 month or multi-year cost model because we dont know is a refreshingly honest admission from someone actually building agentic AI products commercially right now, most vendors arent nearly this candid about the uncertainty

[Ronan_34]

This explains a lot about why enterprise AI contracts keep looking so different from traditional SaaS pricing, you cant just charge a flat monthly seat fee when the actual underlying compute cost per interaction varies this unpredictably

[Nadir Compass]

Agentic systems compounding this problem makes total sense, if a single agent task chains together multiple model calls and sub-tasks, the token consumption and resulting cost can spiral in ways that are genuinely hard to forecast in advance

[Caitlin_69]

Feels like this uncertainty is exactly why so many AI companies default to usage based pricing rather than flat subscriptions, passing the unpredictability directly onto the customer is easier than trying to absorb and average it out themselves

[GlassKnight35]

The businesses recouping massive AI investment through pricing tension is the underlying commercial reality driving basically every AI pricing decision right now, hundreds of billions spent needs to come back somehow and normal customers are the ones who end up footing that bill eventually

[Runtime Dean]

Would love to see more concrete examples of companies that got burned badly by unpredictable token costs after signing a fixed price agentic AI contract, this piece stays pretty theoretical without a specific cautionary tale

[NatureBoy_Dev]

This is a genuinely underappreciated angle in the broader AI economics conversation, most coverage focuses on capex and data center spending but the actual per-interaction pricing unpredictability is just as real a problem for adoption

[GatewayDolphin]

Makes me think third party AI agent vendors are going to increasingly build in large safety margins into their pricing specifically to protect against this unpredictability, which probably means customers end up overpaying on average just so the vendor doesnt take a loss on the unpredictable months

[LordGaz90]

The awkward bit is that token count is not the same thing as value. A short request can trigger a surprisingly expensive chain if the system has to reason, call tools, inspect results, revise its plan, and then produce an answer. Meanwhile, a long document summarisation job might be comparatively predictable.

That makes a simple price-per-token model useful for accounting but not always intuitive for customers. People generally want to buy an outcome, not a pile of tokens.

[MJF86]

The comparison with cloud computing is useful, although AI adds a layer of uncertainty because the workload itself can be generated by the model. Traditional software usually follows instructions written by developers. An agent can decide that another search, another tool call, or another reasoning pass would be useful.

That means providers are pricing not just compute, but a system whose demand can partly decide its own demand. That is a rather unusual billing problem.

[Maya98]

There is also a huge difference between predictable workloads and open-ended ones. A customer asking for 100 document classifications can estimate usage fairly well. An agent asked to investigate an ambiguous problem may take ten steps on one occasion and thirty on another.

That suggests subscription plus usage limits may make more sense for some products than pretending every task has a fixed price. The difficult part is setting the limits so ordinary users do not feel like they are constantly watching a meter.

[Quasar Ruby]

I would also be careful about making prices too complicated in the name of precision. If customers need a calculator, a glossary, and a flowchart just to estimate whether an afternoon of use will cost five pounds or fifty, the pricing has already failed at a usability level.

A simple default with detailed metering underneath seems better. Let ordinary users buy predictable access while power users can inspect every token and tool call if they want to.

[Slay40]

The hardest case may be the occasional user. Someone who uses an AI agent twice a month does not necessarily want another subscription, but they also do not want to learn a pricing spreadsheet before trying it.

Credits could solve part of that. Give the user a visible balance, explain roughly what common tasks consume, and let them buy more when necessary. It is not perfect, but at least the mental model is understandable.

[Leopard10]

The phrase tokenomics makes me laugh a little because it makes every API call sound like a tiny cryptocurrency. But there is a serious issue underneath it: tokens are a provider-facing measurement that does not necessarily map cleanly to what the customer thinks they purchased.

A user thinks they bought a research task, a translation, or an image analysis. The provider has to turn that outcome into measurable compute. Good pricing probably needs to bridge those two perspectives rather than pretending one unit solves everything.

[Kai_37]

I would push back slightly on the idea that token pricing is inherently confusing. Developers already understand it, and transparent usage-based pricing has a real advantage: you can see what caused the bill to increase.

The bigger problem is when the customer is charged for unpredictable internal behaviour without being given meaningful controls. If an agent suddenly decides to make fifty tool calls, there should be a budget, warning, or permission boundary. Otherwise the user is effectively handing over a blank cheque. :/

[Crossing]

There is also a fairness angle. If someone writes concise prompts but receives huge internal reasoning chains, they should not necessarily be punished simply because the model chose an expensive route. Conversely, providers cannot absorb unlimited hidden work just because the customer sees one simple request.

That makes transparency and spending controls more important than finding one magical unit price. Give users predictable ceilings, useful estimates, and a clear explanation of what causes usage to rise, and the exact billing unit becomes much less scary.

[SwiftQuarry]

One thing that gets overlooked is latency. Two tasks might cost roughly the same amount to run but feel completely different if one finishes in ten seconds and the other takes five minutes because it is coordinating several external tools.

For business users, the price of waiting can matter almost as much as the price of inference. That could lead to separate pricing for speed, concurrency, or guaranteed capacity rather than simply charging more tokens.

[David]

There is also a quality problem. A cheaper model may use fewer tokens but produce an answer that needs three rounds of human correction. Suddenly the cheaper inference was not cheaper at all.

For enterprise customers especially, total cost should include review time, failed attempts, tool usage, and reliability. A model that costs twice as much per token but gets the job right on the first attempt can easily be the economical choice.

[Oscar]

Agentic systems make the problem even clearer. Suppose I ask an agent to research five companies and produce a comparison. The visible request might be tiny, but behind the scenes it could involve planning, search calls, several model passes, extraction, checking, and a final synthesis.

Charging only for the original prompt would hide most of the real compute. Charging purely for every internal token could make the bill feel like a taxi meter running while the passenger has no idea where the driver is going.

[ParallelSelf34]

The long-term question is whether AI becomes cheap enough that all this becomes less important for ordinary users. If inference costs fall dramatically, companies may bundle generous usage into subscriptions and reserve detailed metering for heavy commercial workloads.

Until then, the economics are going to remain a balancing act between predictable bills for customers and unpredictable workloads for providers. Agentic systems just make that tension impossible to ignore.

[Olivia87]

The consumer analogy that makes sense to me is electricity rather than software licenses. You can have a predictable base connection, but actual consumption varies. The trouble is that most people can understand a kilowatt-hour reasonably well, whereas tokens are an abstract unit that says very little about how useful the result was.

A pricing model could therefore expose tokens for technical users while giving ordinary customers something like tasks, credits, or completed actions. The backend can still count tokens without making the customer do the accounting.

[Ben55]

One interesting possibility is outcome-based pricing for tightly defined tasks. Something like document extraction could be priced per processed document, while open-ended research stays usage based. The provider can estimate the typical workload for the first category much more reliably.

That would avoid forcing one pricing model onto every AI workload. A chatbot, a coding agent, a transcription service, and a research agent have very different cost structures.

[PlanckLimit36]

I suspect we will eventually see pricing move away from raw tokens for mainstream products. Tokens will remain underneath as the engineering metric, much like CPU cycles and storage operations still exist underneath cloud services, but customers will increasingly buy useful units of work.

The tricky part is defining those units without creating artificial restrictions. An agent that solves a problem in three calls should not be treated identically to one that needs thirty simply because both were labelled a single task.

[NovaBreaker10]

There is a strong argument for giving users an internal usage dashboard as well. Not just a monthly number, but something that says where consumption went: conversation context, tool calls, file processing, agent steps, and so on.

That would make the economics much less mysterious. If I see that one workflow consumed 80 percent of my allowance because it repeatedly searched the same source, I can actually change the workflow instead of just wondering why my credits disappeared.

[NightCrawler81]

There is a neat distinction between cost-based pricing and value-based pricing here. If an AI saves a lawyer three hours, the underlying token cost may be tiny compared with the value delivered. Pricing it strictly according to inference cost could leave money on the table.

But value pricing has its own problem: providers cannot assume every customer gets the same value. A small business and a multinational might use the same feature but have radically different willingness to pay. That is where tiered plans start creeping back in.

[StringTheory32]

A practical model could be hybrid pricing: a predictable monthly allowance for normal usage, explicit charges for heavy workloads, and hard limits on autonomous agents. That would cover the common case without forcing everyone into pure pay-as-you-go billing.

The important part would be making the boundary visible. If an agent is about to consume ten times the normal amount because a task has become complicated, the user should get a chance to stop it.