The AI tools you have not heard of

The conversation about AI in the UK tends to focus on two or three companies.

OpenAI runs ChatGPT. Anthropic runs Claude. Google runs Gemini. Most of the time the rest of the field is a footnote. ChatGPT has been taken on as a more generic term, a bit like Hoover did with vacuum.

There are about thirty large language models available today. Three of them we just mentioned. The other twenty-seven are not often considered, but not because they are not as good. Some are not in it because the conversation has not caught up with them. Some are not in it because they are not based in the United States.

This piece is about the options. What they are. Where they come from. What they cost. And the three ways in.

The other labs

Four are worth naming here, because they are the ones a small firm in the UK is most likely to meet.

Mistral. A French lab, founded in 2023, backed twice by the French government. Its models are open-weights, meaning the model is published and you can download it, run it on your own machines, completely independently, and inspect the weights. Mistral is the most credible European answer to the US labs.

DeepSeek. A Chinese lab, also founded in 2023. Its open-weight model R1, January 2025, and V3, December 2024, are close to the leading models on most tasks, for an order of magnitude less compute. The price is roughly a tenth of the cost.

Qwen. An Alibaba model, the latest family Qwen 3, multilingual. Probably the most-used open-weight model in Asia.

Llama. Meta's open-weight family, first open-weighted in 2023, the latest is Llama 4. Free to download, free to run. Every cloud runs Llama. Every open-source tool starts with it.

There are at least twenty more models in production, mostly outside the US. You will not see them on the news. The list is long. The conversation is short.

There is a shadow to the open-weights idea. A model that anyone can download can be downloaded by anyone, including people you would not want using it. NCSC and the AI safety institutes have written about this. The same feature that gives a small firm in Manchester the answer to its data-residency problem also gives a hostile actor a private model on private hardware. The open-weights choice is a real choice, and it has a real cost on the other side.

How close are they

Two years ago, ChatGPT had a clear edge. Anthropic caught up inside a year. Google shipped Gemini and closed most of the gap by 2025. The open-weight models from Mistral, DeepSeek, Qwen, and Llama are now close to the best on most tasks. The leading US labs still move first on the hardest reasoning work and the new multimodal tools, the open models catch up within months.

For consumers

This is what a small firm actually touches. The chat on the desk, the app on the phone, the AI overview at the top of Google.

OpenAI has ChatGPT on web, iOS, Android. Anthropic has Claude on web, iOS, Android. Google has Gemini on web and mobile, and Google's search engine uses Gemini by default for AI overviews. Mistral has Le Chat. DeepSeek has its own web and mobile app, top of the app stores in January 2025. Qwen has its own chat. Llama is mostly developer and API, no consumer app.

The choice is which one you paste into.

For enterprises

Most enterprises do not call the labs directly. They call the labs through the cloud they already use.

Amazon's cloud hosts the same models inside a firm's own Amazon account, so the data stays where the firm already keeps its data. Microsoft's cloud hosts OpenAI, and Microsoft 365 has Copilot built in, on the desk of every Office user. Google's cloud hosts Gemini, and Google Workspace has Gemini in the side panel.

For small to medium businesses

A small firm is probably on Microsoft 365 or Google Workspace already, which means they are already running Copilot or Gemini in the apps it pays for. The SaaS vendors a small firm uses, the CRM, the accounting package, the HR system, are calling the labs through the cloud environments on the firm's behalf.

If you wire it in

A small firm can also wire the AI in directly. Three ways, with options around them.

The API. Each lab has its own API. An API is the way your software talks to a model, the same way your browser talks to a website. You sign up, you get a key, you pay by the token. The data policy is whatever the lab publishes. Anthropic does not train on API data. OpenAI does not train on API data by default and offers zero data retention for eligible customers. Mistral's paid tier does not train. DeepSeek's terms are worth a careful read.

OpenRouter. A unified API in front of about a hundred models, run by a San Francisco company. One key, one bill, one set of code. The price is the lab's price plus a small fee. The data policy is set per endpoint. OpenRouter's default is zero data retention, the prompts are not stored. LiteLLM is an open-source version you host yourself, with no third party in the middle.

Scaleway. A French cloud, hosts Mistral and a long tail of open-weight models. The data sits in France. The data policy is European by default. The bill is in euros. Scaleway is one of several European clouds, alongside OVHcloud in France, Hetzner in Germany, and the EU regions of Amazon, Microsoft, and Google. Scaleway is the obvious choice for a firm that wants a French bill, a French data policy, and a French data centre. The alternative to all of the above is self-hosted: the firm downloads the model weights and runs them on its own hardware. The data never leaves, the cost is capital not rent, and the operator burden is the firm's.

What it costs

The cost of using the leading US labs is fifteen to thirty dollars per million output tokens, plus a few dollars per million input tokens. A firm pasting a few thousand words a day into the API, the bill is tens of pounds a month. A firm wiring the model into a product, the bill is hundreds to thousands a month.

The open-weight models, hosted on Scaleway or the EU regions of the big three, charge a fraction of that. The trade is the data policy, not the price. Hosting the model inside a firm's own cloud account, on a European cloud, means the data sits where the firm needs it to sit. EU-resident hosting costs more than US-resident, by twenty to forty per cent on equivalent compute, the premium for the data staying in Europe. The choice is sovereignty or cost.

The line

The pros of the wider option set are simple. Sovereignty. Jurisdiction. No training on your paste, on most of them. Cost, especially on the Chinese and open-weight models. The option space itself. The line is what is in the catalogue, not what three labs choose to sell you.

The cons are also simple. The leading models move first and the open models catch up later, the gap much smaller than it was. Operator burden: you pick, you pay, you govern. Governance variation: the data policy is per endpoint and reading the policies is a real cost.

The line for a small firm is the option space. Three labs gave you one choice. Thirty models, three ways in, give you the line. The line is the choice. Tomorrow, the layer above the model.