AI Is Getting Expensive: Should Small Businesses Use Cheaper Models?
- Chris Howell
- 1 day ago
- 10 min read
For the first few years of mainstream AI adoption, many people got used to a simple idea: pay a monthly fee, use the tool, and do not think too much about what each prompt costs.
That period is starting to change. More AI tools are moving towards usage-based pricing, credits, tokens, premium requests, or limits on heavier features. This shift is being driven by the real cost of running large models, especially for tasks that require long context windows, complex reasoning, or repeated automated use. For small businesses, freelancers and local organisations, this matters. AI can still save time, improve decisions and reduce admin work, but the cost of using it carelessly can creep up quickly.
At the same time, cheaper AI models are getting much better. Smaller proprietary models, open-source models, open-weight models and lower-cost systems from companies around the world are now capable enough for many everyday tasks. In some cases, these models are reported to operate at a fraction of the cost of leading proprietary systems, particularly when deployed efficiently or run locally.
So the question is no longer simply “Which AI model is the best?” A better question is: which AI model is good enough for this task, at the right cost and with the right safeguards?
Why AI costs are becoming harder to ignore
AI pricing used to feel predictable. A business could pay for a subscription and assume that most normal usage was included. That model is becoming less common, especially for heavier AI use. Tools that write code, process long documents, run agents, analyse large files or handle repeated automated tasks can consume far more resources than a simple chatbot conversation. Behind the scenes, these tasks require more compute, more memory, and longer processing time, which is why providers are increasingly charging based on usage.
This is where costs can become confusing. A single prompt may seem cheap, but if AI is being used across a team, inside workflows, or through automated tools, those small costs can add up quickly. For example, a staff member using AI for occasional drafting might generate a few thousand tokens per day, costing very little. In contrast, a business using AI to summarise hundreds of documents, generate reports, write code, answer customer queries and run background agents could generate millions of tokens per month. In usage-based systems, that difference can translate into a significant cost gap.
The first use case may be easy to control, but the second needs a budget, a process and some governance. Small businesses do not need to panic, but they do need to stop treating AI as “basically free”.
The rise of cheaper AI models
The other side of the story is more positive. Not every job needs the most powerful frontier AI model. Many everyday tasks can be handled by smaller or cheaper models, especially when the task is clear and low-risk. Recent developments have shown that smaller models can achieve strong performance on structured tasks, particularly when prompts are well-designed. Some open-weight models can be fine-tuned or adapted for specific business needs, improving efficiency further.
There is also a useful middle ground: smaller proprietary models. These are commercial models from major AI providers that are usually faster and cheaper than their flagship models, but still hosted and managed through familiar platforms. For many small businesses, this may be the simplest way to reduce costs without taking on the extra setup and maintenance involved in running open-weight models themselves.
Cheaper models are often suitable for tasks such as drafting first versions of emails or documents, summarising internal notes, turning rough ideas into structured outlines, classifying simple customer feedback, rewriting text in a clearer style, extracting key points from non-sensitive material, or helping with brainstorming and early research. These tasks do not always require the most advanced model available. In many cases, using a high-end model for every small job is like hiring a specialist consultant to alphabetise your stationery cupboard: powerful, but not necessary.
A quick note on model types
Different AI models are designed, priced and controlled in different ways. Some are powerful frontier models, some are smaller hosted models, some are open-weight models, and some can be run locally or through private infrastructure. The names can get confusing, so here is a simple guide.
Model or model family | Typical category | What that means in plain English |
GPT / ChatGPT | Frontier proprietary model | A powerful commercial model family from OpenAI. Usually strong for complex reasoning, writing, coding and broad tasks, but accessed through OpenAI’s tools or API. |
Claude | Frontier proprietary model | A powerful commercial model family from Anthropic, often used for writing, analysis, coding and long-document work. |
Gemini | Frontier proprietary model | Google’s commercial AI model family, often built into Google products and available through Google’s AI services. |
Claude Haiku / Gemini Flash / smaller GPT models | Smaller proprietary models | Commercial models from major providers that are usually cheaper and faster than flagship frontier models. Useful for simpler tasks where you still want a managed, hosted service rather than running a model yourself. |
DeepSeek | Open-weight / lower-cost model family | A Chinese-developed model family that has attracted attention for strong performance and lower operating costs. Suitability depends on hosting, licence, data handling and use case. |
Qwen | Open-weight model family | Alibaba’s model family, with several open-weight versions available. It may be useful for some lower-risk or self-hosted workflows, depending on setup and governance. |
Mistral | Open-weight and commercial model family | A European model company offering both open-weight and hosted commercial models. Often discussed as part of the open AI ecosystem. |
Llama | Open-weight model family | Meta’s model family, commonly used by developers and organisations that want more control over deployment. Licence terms and setup still matter. |
Gemma | Open-weight model family | Google’s open-weight model family, designed for developers and lighter deployment scenarios. |
A frontier model usually means one of the most capable commercial AI systems available at the time. These models are often strong general-purpose tools, but they can be more expensive and are usually accessed through a provider’s platform.
Open-source or open-weight models give users more flexibility, but they are not automatically safer, cheaper or easier. The practical question is still the same: what task are you using it for, what data is involved, where is it hosted, and who is responsible for checking the output?
Where Chinese and open-source models fit in
A lot of recent discussion has focused on Chinese-developed AI models, such as DeepSeek and Qwen, because some of them are powerful, cheap and available as open-weight models. Reports suggest that some of these models have been trained and deployed at significantly lower cost than comparable Western systems, which has raised questions about pricing pressure across the industry.
That does not mean businesses should automatically use them, and it also does not mean businesses should automatically avoid them. The responsible view is more practical than that. The important questions are about how the model is used: where it is hosted, what data is being sent to it, who operates the service, what licence applies, whether the business can control retention and access, whether the model is suitable for the task, and what happens if the answer is wrong.
A model’s country of origin is one factor, but it is not the whole risk assessment. A poorly governed tool from a familiar brand can still create problems, while a cheaper open model used carefully on low-risk tasks may be perfectly sensible. The danger is not “cheap AI” itself, but using it without understanding what it is doing, where the data goes, and whether the output is reliable enough for the job.
Open-source does not always mean simple
The phrase “open-source AI” is often used loosely. Some models are genuinely open-source, while others are better described as open-weight, meaning the model can be downloaded or used more flexibly, but the training data, development process or licence terms may still have limits.
For small businesses, the distinction matters less than the practical reality. What matters is whether you can legally use the model for your purpose, whether you can run it safely, whether someone can maintain it, whether it protects sensitive data, whether it produces good enough results, and whether it actually saves money once setup, support and review time are included.
Running a model locally or on private infrastructure can reduce per-use costs and improve data control, but it may introduce new costs in hardware, setup, maintenance and expertise. A free or cheap model is not automatically cheaper if it takes days to configure, needs specialist hardware, or produces outputs that staff have to constantly fix.
For some organisations, running a model locally may be attractive because prompts and documents can stay on their own machine or private infrastructure. This can be useful for experimentation, internal drafts, or lower-risk workflows where data control matters. However, local AI still needs suitable hardware, setup time, updates, testing and human review. It should be treated as one possible option in the toolkit, not a magic way to make AI free.
When cheaper models make sense
Cheaper models are most useful when the task is low-risk, repetitive, well-defined, easy to check, and does not involve sensitive personal, financial or client data. In these situations, a small business might use a cheaper model to draft internal meeting notes, create first-pass social media ideas, summarise public documents, tidy up rough text, or categorise simple enquiries.
In these cases, the model does not need to be perfect. It needs to be useful, fast, affordable and easy to review. This is where smaller models can shine. They may not be the best at deep reasoning or complex judgement, but they can be more than good enough for everyday productivity work, especially when combined with human oversight.
When frontier models still matter
There are still many cases where using a stronger, more trusted model makes sense. This includes complex analysis, sensitive client work, legal, financial or compliance-heavy tasks, strategic decision support, long documents with important context, work where mistakes could damage trust, and anything involving personal or confidential data.
In these situations, the cheapest option may not be the safest option. A better model, stronger privacy controls, a trusted business account, or a more carefully managed workflow may be worth the extra cost. For client work especially, businesses should be careful about pasting sensitive information into random AI tools just because they are cheap or convenient. Data handling policies, contractual obligations and regulatory requirements should always be considered.
The right answer is not to always use the biggest model, and it is not to always use the cheapest model. The right approach is to match the model to the task.
What is model routing?
One way businesses are controlling AI costs is through model routing.
This simply means sending different tasks to different models. A simple summarisation request might go to a smaller, cheaper model. A complex strategy question, sensitive client task or detailed analysis might go to a stronger frontier model.
For small businesses, this does not need to be complicated. The basic idea is to avoid using the most expensive model for every task. Drafting, tidying text and simple summaries may not need the same level of AI as legal review, financial analysis or high-stakes client work.
Model routing can be handled automatically in more advanced systems, but the principle is useful even without automation: choose the lightest suitable tool for the job, then escalate to a stronger model when the task genuinely needs it.
A simple model-choice framework for small businesses
Before choosing an AI tool or model, it helps to ask a few structured questions. Start by considering the task itself. If it is simple drafting, summarising or organising, a smaller model may be enough. If it involves complex judgement, important decisions or sensitive material, more caution is needed.
Next, consider the data involved. Public information is lower risk, but customer data, employee records, contracts, financial details and private business documents require stronger safeguards, especially under UK data protection regulations. Then think about how accurate the output needs to be. If it is a rough draft, a cheaper model may be fine, but if the output affects a customer, a decision or a payment, review and quality control become much more important.
It is also important to consider how often the task will run. A tool used occasionally is very different from a workflow running hundreds of times a day. Usage-based pricing can become expensive when automation scales, particularly if prompts are long or repeated frequently. Finally, consider who is responsible for checking the output. AI should not remove accountability. Someone still needs to understand the workflow, review the outputs, and know when the model is not suitable.
What UK small businesses should do now
The practical response is not to abandon AI, but to manage it properly. Start by listing where AI is already being used in your business. This should include paid tools, free tools, browser extensions, document assistants, coding assistants, meeting tools and anything staff may be using informally.
Then check the pricing model for each tool. Is it subscription-based, credit-based, token-based, usage-based, or bundled into another product? After that, separate AI tasks into three groups: low-risk tasks where cheaper models may be fine, medium-risk tasks where stronger review is needed, and high-risk tasks where privacy, accuracy and governance matter most.
You may also want to estimate approximate usage levels, especially for automated workflows, to understand where costs could scale unexpectedly. This simple exercise can reveal where money is being wasted, where risks are hiding, and where AI could be used more confidently.
The real lesson: smarter AI, not just cheaper AI
The future of business AI will not be one model doing everything. It will be a mix of tools, models and workflows. Some tasks will use powerful frontier models, some will use smaller proprietary models, some may use open-source or open-weight models, and some may eventually run locally or inside more controlled systems.
For small businesses, the opportunity is clear: AI does not have to be wildly expensive to be useful, but it does need to be chosen carefully. The best AI setup is not the flashiest one. It is the one that helps your business do useful work, protects the right data, controls costs, and keeps humans in charge of important decisions.
If your business is starting to use AI more often, now is a good time to review your tools before the costs and risks grow quietly in the background. Mercia AI helps small businesses think through AI adoption in a practical, responsible and cost-aware way. If you are unsure whether your current AI setup is still the right fit, an AI Readiness Consultation can help you identify where AI makes sense, where it needs guardrails, and where cheaper tools may or may not be appropriate.
