Choosing your model

You have choices: GPT, Claude, and Gemini from the big labs, plus capable open models like Llama, Mistral, DeepSeek, and Qwen you can run yourself. There is no single best model. There is the model that fits this job and this budget, and that changes from one step to the next.

The dimensions that matter

• Capability: how well it reasons on hard, multi-step problems.
• Context window: how much text it can take in at once.
• Speed and latency: how fast it responds, which users feel.
• Cost per token: what you pay, input and output.
• Modality: text only, or also images, audio, documents.
• Tool-use reliability and availability in your region.

Frontier vs small models

Frontier models are the smartest but slowest and most expensive. Many of them can also spend extra tokens thinking before they answer (so-called reasoning models), trading even more speed and cost for better accuracy on hard, multi-step problems. Small models are fast and cheap and completely fine for easy, high-volume steps: classifying a message, extracting fields, writing a short reply. The trick is not picking one model, it is using the right size for each step.

About the context window

A bigger context window lets you stuff in more text, but it is not a free win. It costs more per call, and models can lose detail buried in the middle of a very long context. Retrieving the few most relevant pieces (the RAG lesson) usually beats dumping everything in and hoping.

Open vs closed

Closed models behind an API give you top capability and zero infrastructure work. Open models you host give you control, data privacy, and lower cost at high volume, but you run the servers. Most small businesses start with a closed API and only consider self-hosting when privacy rules or scale make it worth the operational load.

Classify / extract / short reply   ->  small, fast model
Draft with judgement, summarise    ->  mid model
Hard reasoning, tricky tool use    ->  frontier model
Sensitive data, must stay in-house ->  open model you host