What's Behind the Chat Window


When I took a snapshot of the AI model market for our new guide on October 9, Hugging Face, a major public repository for AI models, listed roughly three million model repositories, including variants. OpenRouter, a service that connects software to models from many providers, offered hundreds of model options. These counts measure different things, but both give a sense of the range of choices.

Most people in the workplace meet AI through a text box. They type a question, and an answer arrives in seconds. The box pretty much looks the same, no matter which model sits behind it or which account the person used to sign in.

Those details (the model, the service, and the account) shape important business decisions. Who can see what gets typed into the box, and what can the provider do with it? What happens to the bill when ten people use it every day instead of one? A business can make these choices without anyone stopping to consider them.

In our last issue, I cited the Federal Reserve Banks’ finding that 46% of surveyed small employer firms used AI, but only 7% of those users described it as fully integrated into their business.[1] I also wrote that many owners lack the technical support to evaluate tools and verify vendor claims.

Making Sense of AI Models is a start on that second problem. It’s an illustrated guide to the AI model market, written in plain English for business leaders, and you can access it here. This issue covers the parts I think matter most for owner-led businesses.

The model and the way you access it

“AI” in a business conversation often means a chat app. Underneath it, at least two separate choices are being made.

The first is the model. Models differ in what they’re built to do. General-purpose models handle drafting and summarizing, while others specialize in coding or working with images and audio. Models at the leading edge of capability are often called frontier models. They can carry higher prices, although a smaller, cheaper model may handle many everyday business tasks well.

The second choice is the route: how the model reaches your business. The most familiar route is the provider’s own chat app, such as ChatGPT. Increasingly, though, AI arrives inside software a business already uses. An AI feature in accounting or CRM software may run on a model from one of the major AI labs. In that case, your relationship is with the software vendor, whose contracts and arrangements with its model providers also affect how your data is handled.

Models with downloadable files, known as open-weight models, can also run on equipment you control, subject to their licenses and hardware requirements. That can keep information in-house if the whole setup stays local, but the business becomes responsible for running and maintaining it.

Where your information goes

When someone presses Enter, the request may pass through more than one company. An application receives it, and a model host processes it. Logging services, connected search tools, or other add-ons may also store or receive parts of it. Each can have its own rules.

The plan matters as much as the brand. OpenAI, for example, does not use inputs and outputs from its business plans or API to train its models by default. On personal workspaces, including Free, Plus, and Pro plans, sharing data for training is enabled by default, although users can switch it off.[2] Other providers also distinguish between consumer and business offerings; the guide compares their terms side by side.

Training is only one question. A service that excludes your data from training may still retain conversations or logs. Retention periods vary by product and feature, and zero-retention options can have eligibility and feature limits.[3]

Employees rarely wait for a policy. Verizon’s 2026 Data Breach Investigations Report found that, in the monitored environments it describes, 67% of users accessing AI on corporate devices used non-corporate accounts. It also reported that 45% of employees regularly used AI on corporate devices, compared with 15% in the previous report.[4] These figures are not a representative survey of small businesses, but they illustrate how much AI use can happen outside a company’s chosen accounts. An owner without comparable monitoring may have little sense of how staff are using AI.

Picture an office manager drafting a reply to an unhappy customer. To get the tone right, they paste the customer’s email and account details into a free chat app on their phone. Nobody at the business chose that app or read its terms. The business has still made a decision about where customer information goes.

Any AI service a business relies on should come with clear answers to four questions:

  • Visibility: Which services and people can see what goes in and what comes out?
  • Training: Which vendors may use your content to train or improve their models?
  • Storage: How long is information kept, and where?
  • Sharing: Which other companies receive your information, such as search engines or add-on tools?

For many small businesses, a useful first step is modest: choose the AI accounts the business will support, decide what information may go into them, and make sure people know how to use them. Paying for an account alone does not settle those questions.

What you are paying for

Businesses commonly pay for AI in two ways.

The first is a subscription: a monthly fee per user for a chat app, or an AI add-on to software the business already pays for. Subscriptions are generally predictable, but they can come with usage limits, and heavy users may need higher-priced tiers or extra credits.

The second is usage-based pricing. For text models, the bill usually depends on the amount of input and output, measured in tokens—small units of text. Tools, images, audio, and other features can add separate charges. Owners may encounter this indirectly as credits or overage charges in AI-powered software, or directly when the business builds its own automations.

The model chosen matters a great deal here. As of October 9, 2026, OpenAI’s standard API rates for GPT-6 Astra were 100 times those for GPT-6 Luna for both uncached input and output tokens.[5] That compares equal token volumes, not the cost or quality of completing a particular task.

Why a lower model price may not mean a lower bill

Today’s prices are a poor guide to tomorrow’s. Some are openly promotional: OpenAI’s pricing page says GPT-5.6 Sol’s promotional pricing is available at least through November 21, 2026.[5]

The companies setting these prices are also spending heavily. TechRepublic reports that OpenAI expects about $280 billion in negative cash flow between 2026 and 2030.[6] Some spending builds future capacity and some serves today’s customers, so those projections do not establish the cost of any one product—or tell us what its future price will be.

Where prices go from here is hard to predict, and two arguments point in different directions.

One is that AI keeps getting cheaper at a given level of capability. In February 2025, Sam Altman wrote that the cost of using a given level of AI falls about tenfold every year.[7] Epoch AI’s research also documents steep declines in the price of reaching fixed benchmark performance. Stanford’s 2025 AI Index summarized annual declines ranging from ninefold to 900-fold, depending on the task.[8] These are historical observations, not a guarantee of future prices.

The other argument comes from software pricing. Providers can raise subscription prices even as the underlying technology gets cheaper. Vertice, which tracks software contract spending, reports double-digit annual SaaS price increases in 2026, reaching 16.4% in June.[9] AI is already part of that pattern. When Google announced Gemini’s inclusion in Workspace in January 2025, its US Business Standard annual-plan price rose from $12 to $14 per user per month. Customers who had also been paying for the separate Gemini add-on, however, saw a reduction in their combined price.[10]

Both can happen at once. The price of model access can fall while the price of a tool your business depends on rises. Usage may grow too, as more people and AI agents rely on the tools, so total spending can climb even as each task gets cheaper.

Then there is the cost of leaving. Moving integrations and stored history takes work, and people get used to one tool. That switching cost can give a vendor room to raise prices later.

Open-weight models offer one possible alternative. Where the license permits, a business can run a model itself or choose another host. That gives it options if a provider raises prices, but it does not make a move effortless. Hardware needs vary, larger models can require substantial infrastructure, and someone has to keep the system secure and up to date. Many businesses use a hosting company anyway, bringing back the data questions from the previous section.

Before committing to a provider, I’d ask two questions. What happens to the bill if usage doubles or the price goes up? And what would it take to move somewhere else?

A guide to the choices

I’m noticing a trend in my conversations with business owners and the people who advise them. Most are already using AI, but few have a clear picture of the layers beneath the chat window: which model is doing the work, how it reaches the business, what the bill is really paying for, and where the information goes.

Making Sense of AI Models maps those layers in plain English. It has four parts:

  • The landscape: How big the market is, and what labels such as frontier and open-weight mean.
  • The choices: The ways to access a model, when running AI on your own equipment makes sense, how pricing works, and where your data goes.
  • What’s changing: Routing requests between models, combining several models into one answer, AI built to return a decision instead of prose, and how to read past the headline on the next model launch.
  • A reference: Appendices covering 11 major providers, 19 selected models with relative prices, data terms for 19 services and plans, and a dozen services that provide model access, gateways, or routing, supported by 43 numbered primary-source reference entries.

Model names and prices are a snapshot as of October 9 and will change, but the concepts will hold up much longer. If you advise business owners, feel free to use it in your discussions. When a client asks which AI they should use, it gives you a shared vocabulary to start from.

The wider AI market contains millions of model repositories. A business only needs to understand the tools its people actually use. A good first step is knowing which accounts they’re signed in to and what those accounts allow. If a client or buyer asks how the business handles its data, the answer is easier to give once you’ve looked.

​Access Making Sense of AI Models​

If you’re working through similar questions in your own business, I’d be glad to talk. You can find a time here:

If you’re just interested in learning more about Tether, visit us here:

References

1. Federal Reserve Banks, 2026 Report on Employer Firms: Findings from the 2025 Small Business Credit Survey ​

2. OpenAI, What if I want to keep my history on but disable model training?​

3. OpenAI, Data controls in the API platform​

4. Verizon, 2026 Data Breach Investigations Report: Executive Summary, p. 12​

5. OpenAI, API Pricing. Guide snapshot dated October 9, 2026; cited rates and promotional wording rechecked October 21, 2026.

6. TechRepublic, OpenAI’s Compute Costs Could Burn Through $280 Billion by 2030, September 21, 2026. Reported forecast, not an independently verified outcome.

7. Sam Altman, Three Observations, February 2025​

8. Epoch AI, LLM inference prices have fallen rapidly but unequally across tasks; Stanford HAI, 2025 AI Index: Research and Development, finding 6. The ninefold-to-900-fold summary is historical, not a 2026 forecast.

9. Vertice, SaaS inflation rate, accessed October 21, 2026.

10. Google Workspace, Empowering businesses with AI, January 15, 2025; Google Workspace, payment-plan comparison.

the next shift

Practical ideas on business strategy, AI, and automation to help people navigate disruption and shape what comes next.

Read more from the next shift
The next shift by Sean Wittig: A Second Brain Doesn’t Have to Be Complicated. Start with one recurring problem and the knowledge you already have.

A Second Brain Doesn't Have to Be Complicated Start with one recurring problem and the knowledge you already have. Over the past several weeks, I’ve had many conversations about “second brains.” Some people are building sophisticated systems. Some are tinkering with a few tools. Others have just heard the term and are trying to understand what it means. The term makes it sound like the first brain could use a little backup. Fair enough. It can remember every word of a song you haven’t heard...

The Work We Bring Home: Why I'm Launching Tether Why I’m launching Tether to give small and medium-sized businesses the guidance and ongoing support to improve their operations and put AI and automation to practical use. Under the warm kitchen light, the table was buried in paper. Bank statements, QuickBooks printouts, carbon-copy invoices, and commercial subcontracts covered the surface. My mother’s desktop calculator clacked as she worked through the numbers. “It didn’t come,” my mother...