By Dr. Kevin Shepherdson & Harish Pillay
The cheapest model is not always the cheapest journey - and the most expensive model should not become your default.
Anyone who has booked an airline ticket understands a strange feature of pricing.
Two passengers can sit on the same aircraft, travel to the same destination and arrive at exactly the same time - yet one may have paid S$2,000 while another paid S$12,000.
The economy passenger wants to get from Singapore to London safely and at a reasonable cost. The business-class passenger may be paying for space, flexibility, priority and the ability to work during the journey. Someone flying privately is paying dramatically more, but perhaps because an hour of their time is worth more than the additional cost.
None of these choices is automatically right or wrong. The question is whether the class of travel fits the value of the journey.
AI is beginning to work the same way.
For the first few years of generative AI, many organisations behaved as though every employee should fly business class. They bought access to the best models available and used them for everything: summarising emails, extracting information, preparing presentations, analysing contracts, writing code and increasingly running autonomous agents.
That was understandable when the gap between the strongest and weaker models was large. It is becoming economically difficult to justify today.
The AI market is rapidly separating into different classes of machine intelligence.
1. Cheaper models are becoming capable enough for many routine tasks.
2. Open-weight models can increasingly be run inside an organisation's own infrastructure.
3. At the same time, frontier models such as GPT-6 Astra and Claude Fable 5.1 continue to command premium prices for more complex workloads.
Executives leading AI transformation therefore need to stop asking:
Which AI model is the best?
A more useful question is: What class of intelligence does this particular job justify? That is an allocation problem and increasingly, a management discipline.
The emerging classes of AI
The airline analogy gives leaders a simple way of thinking about model selection.
AI class | Typical model choice | Suitable work | Economic question |
|---|---|---|---|
Economy | Small, specialised, open-weight or low-cost models | Classification, extraction, routing, routine summarisation, simple drafting | Can we perform this reliably at very low cost? |
Premium economy | Strong general-purpose models | Everyday professional work, research assistance, drafting, coding support | Does better capability save enough time or rework to justify the premium? |
Business class | Frontier models such as GPT-6 Astra or Claude Fable 5.1 | Complex reasoning, difficult coding, important analysis, high-value professional work | Is the additional capability valuable enough to justify frontier pricing? |
Private jet | Frontier models with premium speed and complex agentic workflows | Time-critical autonomous work, long-running agents, highly valuable workloads | Is speed or autonomy worth the potentially much higher consumption cost? |
These are not formal industry categories. They are a management metaphor but they illustrate something important.
Not every task deserves a frontier model and not every workflow deserves an agent. And not every token creates the same economic value.
That principle runs through what we call The Economics of AI. The early instinct to use the most powerful model for everything becomes inefficient as the market matures. Organisations will increasingly use portfolios of models, routing work according to accuracy, cost, privacy, latency, quality and risk.
Look at the price difference
The economics become clearer when we put actual numbers against the metaphor.
As of 6 October 2026, OpenAI's standard API pricing for GPT-6 Astra is US$10 per million input tokens and US$50 per million output tokens. GPT-6 Sol costs US$2 input and US$10 output -one-fifth the headline price for the same number of standard tokens.
Anthropic's Claude Fable 5.1 is also priced at US$10 per million input tokens and US$50 per million output tokens, although caching can substantially reduce the effective cost of workloads that repeatedly use the same context.
OpenAI explicitly says that improvements in caching and inference allowed it to cut GPT-6 Sol and Luna API prices by 50% compared with the preceding generation while keeping Astra as the premium model for the most demanding work.
This is exactly what we would expect in a maturing market.
Economy class keeps getting better while Business class does not disappear. The mistake is assuming everybody needs to sit at the front of the aircraft.
Consider a company processing 100,000 relatively straightforward documents every month. If a lower-cost model extracts the required information accurately enough, sending every document to Astra or Fable would be like buying business-class tickets for an entire workforce to attend a routine meeting.
You would certainly get there. But you would be paying for capability you did not need.
But cheaper tokens do not automatically mean cheaper AI
Here is where the airline analogy starts to break down. When you buy an airline ticket, you generally know what the journey will cost before you board.
AI is different. The model may be priced per million tokens, but the user rarely sees everything the system is consuming.
A chatbot reads the prompt and generates an answer.
An agent may interpret an objective, formulate a plan, search the web, retrieve documents, call tools, generate intermediate outputs, evaluate its work, retry failed steps and finally produce something for the user.
One visible instruction can therefore trigger dozens of invisible operations. This is called token amplification: one instruction becomes many operations, operations become multiple model calls, and those calls consume tokens.
Imagine asking an AI: Prepare a competitor analysis and recommend how we should respond.
A simple model might generate an answer from its existing knowledge.
An agentic system might instead identify competitors, search 30 websites, read annual reports, compare products, analyse customer reviews, generate a spreadsheet, critique its first analysis, perform additional research and prepare a presentation.
The second workflow may produce far greater value. It will also consume far more resources. This is why executives should be careful about celebrating falling token prices.
Suppose token prices fall by 50%, but an autonomous workflow consumes five times as many tokens as the process it replaced.
Your AI bill has not fallen.
The real unit of cost is the journey
Airlines do not charge you according to the amount of aviation fuel used by your individual seat. Businesses nevertheless care about the full economics of getting a person from one place to another.
AI leaders need a similar shift in thinking.
The relevant unit should increasingly be cost per successful business outcome, not cost per token.
A model costing US$2 per million tokens can be more expensive than one costing US$10 if the cheaper model:
1. requires repeated attempts;
2. generates more errors;
3. needs more human checking;
4. takes longer;
5. calls more tools;
6. or creates downstream rework.
Conversely, a frontier model may justify its premium if it solves an important problem correctly in one pass.
We called this value per token. Cost per token tells us what machine intelligence costs. It does not tell us whether that intelligence was economically worthwhile. This distinction becomes especially important in AI transformation programmes.
A CFO should not merely ask, "How much are we spending on tokens?"
The better question is: What did those tokens achieve?
Did they reduce cycle time? Lower rework? Improve customer outcomes? Reduce risk? Replace external professional expenditure? Help employees perform work that previously required significantly more time?
A cheap journey to the wrong destination is still wasted money.
Then there are SLMs: owning the aircraft
Small language models and smaller open-weight models introduce another option.
Instead of purchasing intelligence through an API every time it is needed, an organisation can sometimes run a model on hardware it owns.
Alibaba's Qwen3.8-27B, released in August 2026, provides a useful example of this direction. It is a 27-billion-parameter open-weight model released under an Apache 2.0 licence and can be deployed locally using supported inference frameworks and appropriately configured hardware.
For certain workloads, this changes the economics.
The organisation no longer pays a proprietary provider for every input and output token. If it runs the model on its own hardware, it also avoids renting a cloud GPU every time inference occurs.
At first glance, that sounds like free AI. It is not.
There is a total cost of ownership.
The organisation still has to pay for hardware. That hardware depreciates and may become economically obsolete as newer chips arrive. There is electricity, storage, security, software configuration, integration, monitoring, backups, model updates and technical support.
Someone must operate the environment.
Performance also matters. If a local model takes ten times longer than a cloud model, the absence of a token bill does not necessarily make it economical. A highly paid professional waiting for output has a cost too.
There are therefore two very different statements:
"We no longer pay per token."
and
"The AI is free."
They are not the same. Running an SLM locally is closer to owning the aircraft instead of buying tickets.
If you fly once a year, ownership would be absurd. If you operate highly predictable routes continuously, ownership economics become more interesting. The same applies to local inference.
High-volume, predictable workloads -especially those involving sensitive data or requiring greater control -may justify the capital and operating cost. Low-volume or rapidly changing workloads may remain cheaper through an API.
That decision should be based on TCO (Total Cost of Ownership), not on the seductive phrase "free model".
For a first level TCO consideration of locally run models, items like hardware (and depreciation costs), power, system administrators, redundancy (BCP) have to be balanced against sovereignty of use. TCO of cloud provider solutions would include token costs, connectivity costs and the business risks of system access denial.
Why the AI providers are changing their tickets too
Pricing changes at OpenAI also show why the market is becoming more segmented.
As of October 2026, ChatGPT Pro is offered at US$100, US$200 and US$500 per month. The US$500 plan includes Astra Ultrafast, while OpenAI confirms that new US$200 subscriptions now receive a lower usage allowance than the earlier Pro 200 plan.
This should not automatically be interpreted as consumers being "ripped off".
There is another economic explanation. Heavy AI users are expensive customers.
Traditional SaaS economics encouraged vendors to want customers to use software as much as possible because serving one more click or spreadsheet had very low marginal cost.
Generative AI changes that relationship. Every additional inference consumes compute and every longer context consumes more. Every additional agentic loop consumes more again.
Providers therefore have an economic incentive to differentiate between ordinary users and people consuming enormous amounts of frontier compute.
Airlines discovered this decades ago. One ticket price cannot economically serve every customer equally.
AI providers are now discovering something similar.
The executive model-allocation test
For leaders undergoing AI transformation, I would therefore introduce a simple decision before approving any major AI workflow.
Ask six questions:
1. What is the economic value of the task? A board-level acquisition analysis deserves more intelligence than reformatting an internal memo.
2. What happens if the model is wrong? Higher-risk work may justify better models and stronger human review.
3. What is the cheapest model that performs the task reliably? Do not pay for intelligence you do not use.
4. What does the whole workflow consume? Include context, retries, agents, tool calls and hidden intermediate work.
5. What human cost remains? Review, correction and escalation belong in the calculation.
6. Would local inference reduce total cost of ownership? Include hardware, depreciation, electricity, integration, maintenance, security and technical support -not merely avoided API tokens.
This turns model selection from an IT purchasing decision into an AI resource allocation decision.
Do not fly business class everywhere
There will always be organisations willing to pay for frontier intelligence. And they may be absolutely right to do so.
If a frontier model helps an engineering team solve a problem worth millions of dollars, the difference between US$2 and US$10 per million input tokens is almost irrelevant.
If an inexpensive model is classifying routine documents millions of times a month, the same price difference becomes material.
Context determines value. So the economic discipline for AI is surprisingly similar to buying an airline ticket:
Do not fly business class for every journey.
But equally:
Do not buy the cheapest seat when the cost of arriving late - or at the wrong destination - is much higher than the ticket.
The winners in AI transformation will not necessarily be the companies with the biggest AI budgets or the most powerful models. They will be the organisations that learn how to allocate different classes of machine intelligence to the work where each creates the most value.
That is where token management becomes something much larger than controlling an AI bill. It becomes the economics of AI.
Sources: OpenAI Developers, Anthropic, GitHub, OpenAI Help Center, OpenAI
