AI models
A plain-English guide to the major AI providers
Who builds the models, what each is genuinely better at, and why the right answer for your business is usually more than one of them.
The short version. A handful of companies build the large AI models that everything else runs on. They are close in capability and they leapfrog each other every few months. The differences that matter to a business are rarely about raw intelligence. They are about cost per job, how much text the model can hold at once, where your data is processed, and how well it follows instructions without improvising.
Jargon, translated
The eight words you will hear
You do not need these to work with us. They are here so nobody can use vocabulary to make a simple thing sound expensive.
Model
The AI itself: the thing that reads text and produces text. Claude, GPT and Gemini are models. Everything else on this page is about how you use one.
Token
How usage is billed. A token is roughly three quarters of a word. You pay per token going in and per token coming out, which is why a system that reads a 200-page contract costs more to run than one that reads an email.
Context window
How much the model can hold in mind at once. A large window means it can read an entire contract set before answering, instead of being fed fragments.
Hallucination
When a model states something false with complete confidence. The fix is to make it answer only from your own documents and show where each answer came from, and to hand the question to a person when it is unsure.
RAG (retrieval-augmented generation)
Jargon for “look it up before answering”. Your documents are indexed; the relevant passages are retrieved and handed to the model along with the question. It is how an assistant answers about your business rather than about the internet.
Fine-tuning
Further training a model on your own examples so it adopts a style or a narrow skill. Useful less often than people expect, because retrieval and clear instructions solve most problems more cheaply, and are easier to change later.
Agent
A model given tools and permission to use them (read a record, send a draft, update a row), usually over several steps toward a goal. This is where most business value is today, and where controls matter most.
MCP
An open standard for connecting AI to tools and data, so an integration built once can be reused across models. It is plumbing, but it is part of why swapping providers is now a configuration change rather than a rewrite.
The landscape
Who the major providers are
Specific model versions move too fast to publish on a website. What follows is the shape of each provider, which changes far more slowly.
Anthropic
Claude family · closed
Strong at long, careful reasoning over messy real-world context, at following detailed instructions without drifting, and at using software tools reliably. Those are the qualities that matter most when a system is allowed to touch your business records. Widely used for agents that act rather than only chat.
OpenAI
GPT family · closed
The broadest ecosystem and the most familiar to your team, because ChatGPT is where most people met AI. Very strong at structured output, images and voice, with a wide range of model sizes, which makes it the economical choice for high-volume, well-defined steps.
Gemini family · closed
Very large context windows, useful when a job means reading an entire contract set or a year of transcripts at once, plus native proximity to Google Workspace and Google Cloud. Often the pragmatic choice when your business already lives in Google.
Meta
Llama family · open weights
Published openly, so the model can run on infrastructure you control. Generally a step behind the best closed models on the hardest reasoning, but more than capable for classification, extraction and routing, and nothing leaves your environment.
Mistral
Open weights & hosted · European
A European provider with both open-weight and hosted models, efficient at small sizes. Frequently the answer when data residency inside the EU is a hard requirement rather than a preference.
The open ecosystem
Hugging Face · Ollama · specialist models
Thousands of smaller models, many tuned for one narrow job: transcription, translation, classification, extraction. Cheap, fast and self-hostable. A well-chosen small model often beats a large one on a narrow task, at a fraction of the cost.
The one real fork
Closed models or open models
Nearly every other decision follows from this one, and it is a business decision rather than a technical one.
| Closed / hosted | Open weights | |
|---|---|---|
| Examples | Claude, GPT, Gemini | Llama, Mistral, specialist models |
| Where it runs | The provider's servers, reached over an API. | Your cloud, or your own hardware. |
| Capability | Highest available, particularly on hard reasoning. | Very good and closing the gap, and excellent on narrow tasks. |
| Cost shape | Pay per use. No upfront cost, scales with volume. | Pay for infrastructure. Cheaper at high volume, more to set up. |
| Your data | Leaves your network. Business terms exclude training on your inputs. | Never leaves your environment. |
| Best for | Almost every business starting out. | Regulated data, strict residency rules, or very high volume. |
Most setups end up running closed models for the hard work and open models for the high-volume or sensitive parts. It is not either/or.
Our approach
One model per task, not one model per company
Committing a whole business to a single provider is a bet that one company will stay ahead forever. None of them has managed that yet.
We build so the model is a replaceable part. Each step of a job runs on whichever model is best for that step, and swapping one out later is a configuration change rather than a rewrite.
- Cheap, fast models for sorting, extracting and classifying
- The strongest models for judgement, drafting and multi-step work
- Self-hosted models where the data must not leave your environment
- Regular re-testing, because the ranking genuinely does change
What this saves you
- Cost. Routing simple steps to small models is routinely an order of magnitude cheaper than sending everything to the largest one.
- Risk. If a provider raises prices, changes terms or has an outage, your business is not stranded.
- Time. When a better model lands, you get it without a rebuild.
Which of your problems is worth fixing first?
That is the useful question, and the model choice follows from it. We will map both with you.