Loading the Elevenlabs Text to Speech AudioNative Player…


AI is getting cheaper, more capable and considerably more complicated to buy. Behind every chatbot, automated workflow and enterprise AI subscription sits an invisible economy of tokens, processing power and usage limits. Understanding how that economy works could make the difference between an AI investment delivering genuine commercial value and another expensive software subscription nobody quite understands.

The AI Industry Has Discovered Something Rather Familiar: Charging for Consumption

For years, businesses have been buying AI subscriptions with roughly the same level of financial scrutiny they apply to office coffee.

Twenty pounds here, a hundred pounds there, another premium licence for someone who insists they need the most intelligent model available to write their weekly update.

Individually, none of it looks particularly alarming.

Then the finance director discovers 200 AI licences, several development teams running API integrations, a collection of autonomous agents and an invoice nobody can adequately explain.

Welcome to AI tokenomics.

As someone who works with businesses on AI strategy and practical implementation, I find the economics of AI increasingly interesting. The technology gets most of the attention, but the real commercial questions are often rather less glamorous.

How much processing power does a business actually need? Which tasks justify an expensive reasoning model? And what happens when an apparently inexpensive AI assistant starts running thousands of automated operations?

These questions matter because AI is moving from individual experimentation into business infrastructure.

The challenge is no longer simply getting access to intelligence. It’s understanding how much of it you’re buying and whether you’re using it sensibly.

What Are AI Tokens, and Why Do They Cost Money?

A token is a small unit of information processed by an AI model.

For English text, one token represents approximately four characters or three-quarters of a word, although the precise number varies according to language, formatting and tokenizer.

Every API interaction generally involves two categories.

Input tokens include instructions, questions, previous conversation, retrieved information and documents supplied to the model.

Output tokens include generated responses and, where applicable, internal reasoning tokens that may not appear in the final answer.

Providers commonly charge more for output than input because generating text involves sequential computation.

Consider a hypothetical model charging $2 per million input tokens and $10 per million output tokens.

Usage Tokens Cost
Input 10,000 $0.02
Output 2,000 $0.02
Total 12,000 $0.04

Four cents sounds wonderfully economical.

And it is, until you repeat that operation thousands of times, add extensive documents, enable deeper reasoning or allow multiple agents to interact.

The cost of one interaction rarely causes trouble.

The cost of poorly designed repetition can.

Why Do Long AI Conversations Consume More Tokens?

One of the least understood aspects of AI is how conversational context works.

In a typical API implementation, the application supplies relevant conversation history with each new request so the model can maintain continuity.

Imagine asking an assistant to analyse a 100-page business document.

The first question requires the document and your instructions.

The second may require the same document, the original question, the first answer and your new question.

By the twentieth exchange, considerably more information may be processed than at the beginning.

If an application continually resubmits the entire conversation, cumulative input consumption can grow approximately quadratically with the number of turns.

That doesn’t mean every AI product operates this way. Caching, retrieval, summarisation and context-management techniques can reduce unnecessary processing.

But the commercial principle remains.

The longer and more information-heavy a workflow becomes, the more important its architecture becomes.

A poorly configured assistant can spend more time repeatedly reading information than doing anything useful with it.

Which sounds suspiciously like certain management meetings.

For technical detail, see the OpenAI prompt caching documentation.

Why Are There So Many Different AI Models?

OpenAI, Anthropic, Google and other providers offer different combinations of lightweight, general-purpose and advanced reasoning models.

The underlying logic is relatively straightforward.

Not every business task needs the same level of intelligence.

You wouldn’t ask your most expensive strategic consultant to alphabetise the filing cabinet. Yet businesses regularly apply the AI equivalent by assigning sophisticated reasoning models to basic classification and administrative work.

There are three useful categories.

Lightweight models

Fast, economical models suited to repetitive tasks such as categorising enquiries, extracting information, tagging documents and identifying customer intent.

Balanced models

General-purpose systems suitable for document analysis, coding, content development, customer service and business automation.

Advanced reasoning models

More capable systems designed for demanding analytical tasks, complex software development and multi-stage problem-solving.

The objective isn’t simply to buy the cheapest intelligence.

It’s to buy the least expensive intelligence that can reliably complete the task to the required standard.

That’s a rather more useful procurement principle than automatically selecting whichever model has the most impressive launch video.

AI Token Costs by Model: November 2026 Comparison

The differences between model tiers become particularly clear when examining their proposed API pricing.

The following tables reproduce figures from the research supplied for this article.

Important pricing note: These figures have not been independently verified against live vendor price lists. Some model names, availability claims and rates require confirmation. They are included as a research comparison, not as established November 2026 prices or a procurement quotation.

All figures are in US dollars per million tokens.

OpenAI model pricing

Model Input Cached input Output
GPT-6 Astra $10.00 $1.00 $50.00
GPT-6.1 Sol / GPT-6 Sol $2.00 $0.10 $10.00
GPT-6.1 Terra $1.00 $0.05 $5.00
GPT-6 Luna $0.10 $0.01 $0.50
GPT-5.6 Sol $5.00 $0.50 $30.00
GPT-5.6 Terra $2.00 $0.20 $12.00
GPT-5.6 Luna $0.20 $0.02 $1.20

Official reference: OpenAI API Pricing

Anthropic Claude model pricing

Model Input Cached input Output
Claude Fable 5.1 $10.00 $0.25 $50.00
Claude Opus 5.5 $4.00 $0.20 $20.00
Claude Sonnet 5.5 $2.00 $0.10–$0.20 $10.00
Claude Haiku 5.5 $0.10 $0.01 $0.50
Claude Haiku 4.5 $1.00 $0.10 $5.00

Official reference: Anthropic API Pricing

Google Gemini model pricing

Model Input Cached input Output
Gemini 4 Argon $2.00 $0.10 $10.00
Gemini 3.1 Pro $2.00 $0.20 $12.00
Gemini 3.8 / 3.6 Flash $0.75 $0.075 $3.75
Gemini 3.1 Flash-Lite $0.25 $0.025 $1.50

Official reference: Google Gemini API Pricing

The supplied research also describes promotional pricing and future increases for certain Gemini models. Those dates and rates require confirmation.

xAI Grok model pricing

Model Input Cached input Output
Grok 4.7 $2.00 Not specified $6.00
Grok 3 $3.00 Not specified $15.00
Grok 3 Mini $0.30 Not specified $0.50

Official reference: xAI Developer Documentation

Methodology: The tables use the model names and prices in the supplied research, without conversion into sterling. They do not include regional taxes, tool charges, long-context surcharges, batch discounts or separate reasoning overhead. Cached-input pricing may involve different eligibility rules across providers.

What Do Those Token Prices Actually Mean for a Business?

A price per million tokens is useful, but it isn’t particularly intuitive.

Let’s translate it into a business example.

Imagine an organisation processing 10 million input tokens and generating 2 million output tokens each month.

Using the research-supplied rates, and assuming no caching, surcharges or additional processing costs, the calculation is:

Monthly cost = (10 × input price per million) + (2 × output price per million)

Model Input cost Output cost Total
GPT-6 Luna $1 $1 $2
GPT-6.1 Terra $10 $10 $20
GPT-6.1 Sol $20 $20 $40
Claude Sonnet 5.5 $20 $20 $40
Claude Opus 5.5 $40 $40 $80
GPT-6 Astra $100 $100 $200

Illustrative calculations based on unverified research prices, not confirmed vendor quotations.

The difference between the lowest and highest examples is a factor of 100.

That doesn’t mean every model can perform the same task to the same standard.

It does mean businesses should stop treating model selection as a technical preference with no commercial consequences.

The important question isn’t whether a model costs $2 or $200.

It’s whether the $200 model delivers an outcome that justifies the difference.

What I’ve Learnt From Working With Businesses on AI Implementation

Over more than three decades working in business strategy, marketing and digital technology, I’ve seen plenty of technology investments where enthusiasm arrived considerably earlier than commercial discipline.

AI is no exception.

Through my work at Emotio, including AI workshops and operational discovery with hospitality businesses, I’ve been exploring how AI can support areas such as guest intelligence, internal knowledge access, operational reporting and workflow automation.

These are practical business problems, not abstract demonstrations of what a language model might theoretically achieve.

One of the most useful distinctions is between a task that needs genuine reasoning and one that primarily needs reliable access to the right information.

Take an operational assistant preparing a daily briefing.

Its job might involve retrieving relevant information from several systems, identifying exceptions and presenting a useful summary.

The expensive part doesn’t necessarily need to be the language model. The difficult work may be connecting systems, maintaining accurate data, managing permissions and ensuring the output can be trusted.

A more powerful model won’t compensate for unreliable source information.

Similarly, an internal knowledge assistant might answer hundreds of routine questions using approved company documentation. Sending every enquiry to a premium reasoning model could add cost without improving the answer.

The design principle I favour is simple: start with the business outcome, establish the information and controls required, then choose the model.

This is an implementation perspective, not a claim that any particular client deployment has achieved a measured cost saving.

And it highlights something that often gets overlooked.

The model may be the cleverest part of an AI system, but it isn’t necessarily the hardest part to get right.

How Model Routing Can Reduce Enterprise AI Costs

A sensible AI architecture doesn’t automatically send every request to the same model.

Instead, it can route tasks according to complexity.

Imagine a system handling 10,000 customer enquiries each month.

A lightweight model identifies the enquiry type.

A balanced model handles questions requiring additional analysis or customer-specific information.

A more advanced model, or a human employee, deals with exceptional cases.

This is commonly described as model routing or model cascading.

The potential savings come from using expensive reasoning selectively.

However, routing introduces additional complexity. A lightweight model must recognise when it is out of its depth, and repeated attempts can quickly undermine the savings.

A cheap answer that requires three corrections isn’t necessarily cheaper.

Businesses should measure successful resolution rates, escalation, response time, accuracy and total cost per completed task.

ChatGPT Business, Claude Pro and Enterprise Licences: What Are You Buying?

One of the biggest sources of confusion is the difference between an AI subscription and direct API consumption.

A business subscription generally provides access to an application, its interface, administration, collaboration features and a defined level of usage.

An API provides developers with programmable model access, commonly charged according to consumption.

These are not interchangeable.

A ChatGPT Business subscription, for example, shouldn’t be treated as a prepaid allowance of API tokens.

Likewise, a premium individual subscription doesn’t necessarily provide unrestricted access to every model, tool or autonomous workflow.

Usage limits can depend on the selected model, workload complexity, tools and provider policies.

For most organisations, the sensible distinction is:

Business requirement Suitable approach
Everyday writing and research Managed AI subscription
Team collaboration and governance Business or enterprise workspace
Automated CRM processing API integration
High-volume document classification API with appropriate model and batch processing
Complex recurring agents Metered API or supported enterprise agent platform

Official product references include the ChatGPT Business overview and Anthropic’s Claude pricing.

The subscription buys a working environment.

The API buys programmable processing.

Both have a place, but confusing them makes capacity planning unnecessarily painful.

The Hidden Costs of AI Agents and Extended Reasoning

AI agents introduce another financial consideration.

A conventional chatbot might answer one question.

An agent could interpret the request, search internal documents, query a CRM, call another model, analyse the results, correct an error and generate a report.

What appears to be one user request may involve dozens of underlying operations.

Extended reasoning adds another variable. Some models consume additional output tokens while working through difficult problems, even though the user only sees the final answer.

That can be valuable for demanding analysis.

It’s less valuable when the system spends substantial computational effort deciding whether an incoming email belongs in sales or accounts.

Businesses should monitor model calls, reasoning consumption, tool usage, retries and failed executions.

Without those controls, autonomous systems can become impressively efficient at spending money.

What Does AI Tokenomics Mean for Return on Investment?

The most important measure isn’t cost per token.

It’s cost per successful outcome.

Klarna’s widely reported early AI customer service deployment handled approximately 2.3 million conversations and was described as performing work equivalent to 700 full-time agents. The company initially projected substantial financial benefits.

Those figures generated considerable interest, but they should be understood as company-reported outcomes and projections, not a universal benchmark for AI automation.

The lesson isn’t that every business can replace its customer service department with a chatbot.

It’s that AI can change the economics of service delivery when automation is accurate, appropriate and properly supervised.

A useful financial framework is:

Net AI value = measurable financial benefit – total AI operating cost

Costs should include subscriptions, API consumption, integration, monitoring, human review, maintenance and errors.

Benefits might include reduced handling time, increased conversion, faster delivery, additional capacity or genuinely avoidable expenditure.

Saving an employee 30 minutes doesn’t automatically create 30 minutes of financial value.

The organisation needs to establish what happens to that time.

Otherwise, we’ve simply created a faster route to the next meeting.

How Can Businesses Control AI Spending?

The strongest cost controls are often architectural rather than contractual.

Use prompt caching intelligently. Repeated instructions and reference material may qualify for discounted processing, depending on provider rules.

Process non-urgent work in batches. Document extraction, classification and scheduled reporting may benefit from asynchronous processing.

Control context growth. Don’t repeatedly supply entire documents when targeted retrieval or smaller relevant extracts will do.

Set budgets and alerts. Attribute consumption to departments, applications and projects. Where provider alerts aren’t hard limits, enforce application-level controls.

Measure quality alongside cost. A cheap model producing unreliable answers can become an expensive business decision.

Bottom Line: The goal isn’t minimum token consumption. It’s the lowest sustainable cost for a reliable, useful result.

What Should Business Leaders Do Next?

For most organisations, AI tokenomics doesn’t require the board to become expert in GPU architecture.

It requires a few commercially sensible decisions.

First, understand where AI is already being used. Identify subscriptions, integrations, automated workflows and the departments responsible for them.

Second, classify those activities by business value and technical complexity. Routine processing and strategic analysis shouldn’t automatically receive identical resources.

Third, establish ownership of AI expenditure. Finance, technology and operational leadership need a shared view of consumption, quality and return.

Finally, review the economics regularly.

Model prices change, capabilities improve and workflows that were expensive six months ago may become considerably cheaper.

The reverse can also happen when usage expands without controls.

AI procurement should be treated as an ongoing operational discipline rather than a one-off software purchase.

The Real Cost of AI Isn’t Always on the Invoice

The AI industry is moving towards a market where intelligence is available at different prices, speeds and levels of capability.

That creates an enormous commercial opportunity.

But it also demands more sophisticated buying decisions.

Businesses that understand tokens, context, model selection and capacity management will be better placed to build useful AI systems without allowing operating costs to drift.

Those that don’t may discover that their most expensive AI problem isn’t the technology itself.

It’s paying premium prices for ordinary work, automating processes nobody needed and measuring success by how many people have logged in.

AI tokenomics isn’t ultimately about counting tokens.

It’s about understanding the economics of useful intelligence.

And that’s a conversation worth having before the next invoice arrives.

Sources and Methodology

This article combines the author’s professional perspective with supplied technical and market research.

Primary documentation for model pricing and API mechanics:

These links identify the appropriate primary sources for verification. Do not interpret them as confirmation of the research-supplied prices shown above.

Pricing methodology: All rates are expressed in USD per million tokens. The illustrative calculations assume fixed input and output volumes, with no caching, reasoning overhead, tool charges, batch discounts, taxes or additional surcharges.

AIG Agents
What is an AI Agent?AIAI Insights

What is an AI Agent?

Damon SegalDamon SegalMarch 25, 2025
AI Hardware
The Interplay of Hardware and Energy in Advancing Artificial IntelligenceAIPhysical AITech

The Interplay of Hardware and Energy in Advancing Artificial Intelligence

Damon SegalDamon SegalJanuary 31, 2025
AI News 31 January 2025
This Week in AI, AGI, and ASI: The Latest DevelopmentsAI News

This Week in AI, AGI, and ASI: The Latest Developments

Damon SegalDamon SegalFebruary 1, 2025
The owner of this website has made a commitment to accessibility and inclusion, please report any problems that you encounter using the contact form on this website. This site uses the WP ADA Compliance Check plugin to enhance accessibility.