Lesson 3: Choosing Between Models
By the end of this lesson, you will be able to:
- Distinguish the model layer from the platform layer, and explain why they are separate decisions (K9)
- Recognise when a platform chooses the model for you and when the choice is yours to make (K9, K25)
- Compare the leading model families across cost, privacy, and capability (K9, S25)
- Apply a structured set of considerations — capability and fit, tone, ecosystem, cost, privacy, and organisational policy — to choose a model when the choice is yours (K9, S25)
- Use a model-comparison reference to reason about cost and privacy trade-offs before any real data touches an LLM (K9, S25, K2)
- Decide which model powers your Evidence Companion, identify which of the three model-choice situations it is in, and justify that choice against the factors that matter for its task (K9, S25)
Two layers: the platform and the model
The previous lesson showed you how an LLM works and why Claude, GPT, and Gemini behave differently even when their headline benchmarks look similar. This lesson is about the decision that follows from that: when you build automation, which model does the actual reasoning, and is that choice even yours to make?
It helps to see two layers. A platform is where you build and run the workflow. A model is the LLM that does the actual thinking inside it. These are separate layers. The same platform can often run different models, and the same model can be reached through many different platforms. Unit 2 looks closely at the platform layer. This lesson is about the model layer that sits inside it.
🔑 Platform layer and model layer: The platform is where the workflow is built and run. The model is the LLM that performs the reasoning steps inside the workflow. They are separate decisions, even though some platforms bundle them together so tightly that the model choice is hidden from you.
Many never notice the model layer, because for a great deal of no-code work the platform makes this decision for them. Three situations are worth recognising.
The platform decides for you. When you use AI Builder inside Power Automate for a standard task such as sentiment classification, Microsoft selects the underlying model. You do not choose it, and for many routine tasks you do not need to. The platform has made a reasonable default choice on your behalf.
The platform lets you choose. In n8n's AI Agent node, or in Copilot Studio, you can select which model the agent uses. Copilot Studio, for example, can run on OpenAI's GPT models and, since January 2026, on Anthropic's Claude. Note that Claude is not the default in Microsoft and is not available out of the box: GPT remains the default model for new agents, and an administrator must first enable Anthropic models in the Microsoft 365 Admin Center before anyone in the organisation can select them. Even so, where the choice is available the model is a real decision with real consequences for quality, cost, and data handling.
n8n AI model selection via the AI Agent node. Reference: freeCodeCamp — How to Build AI Workflows with n8n.
The platform is the model. In an AI-native productivity tool such as Claude Cowork, the model and the platform are effectively the same choice, because the product is built around one model family.
Reasoning about the model layer
When the model choice is yours, you need a way to compare the options. A model is not just "better" or "worse" in the abstract; it has a capability profile, a cost profile, and a data-handling profile, and the right choice depends on which of those matters most for your task.
To make that comparison concrete, the reference below sets the five leading model families side by side: OpenAI's GPT, Anthropic's Claude, Google's Gemini, Microsoft Copilot, and Meta's Llama. For each, it summarises the free and paid tiers along with their cost, usage limits, and privacy terms. Use it as a starting map, not a final verdict, and always confirm the current specifics against each provider's own documentation, since these details change often.
LLM comparison: the five leading model families across cost, usage, and privacy. Correct as of June 2026 — always confirm current specifics against each provider's own documentation. 📄 Download the PDF
How to choose a model
The comparison table tells you what each family offers. It does not tell you how to decide between them — and there is one fact about the current landscape that changes how you should weigh that decision.
At the frontier, the models have largely caught up with each other. Stanford's 2026 AI Index found the leading models clustered within about 25 Elo points on the public leaderboards: the top systems now sit within roughly three percent of each other, and on many professional benchmarks the top fifteen models are separated by as little as three percentage points. The race is being run at the very top, and no single model dominates.
🔑 Elo points: a rating system borrowed from chess for ranking competitors by relative strength. What matters is the gap between two ratings, not the absolute number. A difference of a few hundred points means one competitor clearly beats the other; a gap of only about 25 points is effectively a dead heat — the two win against each other almost equally often. So "the leading models are within 25 Elo points" is another way of saying they are neck and neck.
That has a practical consequence, especially for your own personal development. If the leading models are this close, then for most everyday work you are no longer choosing a "best" model — you are choosing a fit. And rather than constantly model-shopping for whichever system topped the leaderboard this month, you will get a lot of more values choosing one capable model and getting genuinely deep with it: learning its quirks, where it is strong, and how to prompt it well. Fluency in one good model beats shallow familiarity with five.
Did you know?
In June 2026 Anthropic released Claude Fable 5 with the strongest safeguards it had ever applied. Within three days researchers found a way to prompt it into flagging software vulnerabilities, and the model was withdrawn worldwide for about two weeks before redeploying with a stronger safety filter. The revealing detail for model choice: Anthropic found that several older and rival models could surface the same vulnerabilities. That is convergence made visible — capability at the frontier is now so close that no single model holds a unique edge for long, which is exactly why "fit" matters more than chasing the top of a leaderboard. Anthropic — Redeploying Claude Fable 5.
When the choice is genuinely yours, work through these factors — starting with the everyday ones and ending with the formal gate.
1. Capability and fit. Start with the task, not the brand. Each family has a stated strength: a versatile all-rounder, a leader for coding and writing, a strong fit for one productivity suite, a native option for another, an open-weight option for full data control. Match that to what you actually need — including the practical fit of modality (text, image, voice, vision) and context window — rather than to the headline benchmark.
2. Tone and style. Each model has a recognisable "house voice." One tends to be careful and nuanced, another punchy and concise, another terse and fact-first. Because raw capability has converged, this is often the deciding factor for writing and drafting, and it is subjective: the only reliable test is to run the same prompt through two or three models and see which voice fits your work.
3. Ecosystem fit. You may already have a capable model where your work lives — Copilot inside Microsoft 365, Gemini inside Google Workspace. A model that is already integrated into your daily tools removes friction, and that convenience can outweigh a two-percent benchmark edge.
4. Cost. Notice how differently the families charge. Some charge per seat per month. Some charge per million tokens. Some are free to self-host but carry infrastructure cost. The cheapest option at low volume is not always the cheapest at scale.
5. Data handling and privacy. Notice the consistent pattern the comparison draws out: consumer chat tiers generally train on your conversations by default, with an opt-out, while business, enterprise, and API tiers generally do not train on your data and carry formal security commitments. For any workplace automation touching real organisational data, the tier you choose matters as much as the model family.
If you use a personal or free account, "train on your conversations" is usually switched on by default — so go and check it, because most people never realise it is there. In Claude, open Settings → Privacy and turn off the model-training / "help improve Claude" toggle if you do not want your chats used for training. The equivalent lives under Settings → Data controls in ChatGPT and under Gemini Apps Activity in Gemini. The labels move around, but the principle is constant: find the training setting and make a deliberate choice rather than accepting the default — and never rely on a consumer tier's toggle for genuinely sensitive work material, which belongs on a sanctioned business or enterprise tier.
6. Your organisation's AI policy — the final gate. Even a model that wins on every factor above is the wrong choice if your organisation has not sanctioned it for the data you are using. Reaching for an unapproved tool ("shadow AI") is itself a data-handling decision, made without authority. Before you commit, confirm the model and the tier are approved for your use case.
Optional — go deeper on model tiers
Take your own project's data type and find it in the model comparison. If your project touches personal data, customer data, or anything commercially sensitive, which tiers would you rule out immediately, and why? Write down your reasoning and bring it to your coach.
Choose the model that powers your Evidence Companion
In Lesson 1 you created your Evidence Companion and wrote its first system prompt. That prompt shapes how it behaves; this lesson is about the engine underneath it — which model does the reasoning. Before you refine the Companion further, be deliberate about this choice. The first step is to work out whether the choice is even yours, using the three situations from the start of this lesson.
Look at the tool you built your Companion in:
- If you built it in Claude Cowork, you are in "the platform is the model" situation — the reasoning is done by Claude. You may still get to pick a tier (a faster everyday model versus a slower, stronger one), but the family is fixed by the product.
- If you built a Copilot agent, you are in "the platform lets you choose." Copilot agents run on GPT by default, and since January 2026 can run on Claude where an administrator has enabled Anthropic models. The model is a real decision you can revisit.
Either way, be clear about the Companion's job: drafting and structuring your evidence from rough notes. That is a writing-and-reasoning task, not high-volume classification, and you will be pasting real notes about your work into it. So the factors that dominate here are capability and fit, tone and style (you will read a lot of its prose), and data handling plus your organisation's policy (the notes are real workplace material). Cost matters far less at one-person, low-volume use than it will at scale.
[DECIDE IT — 10 minutes] Choose your Companion's model and record why
Evidence Companion — Model Choice
Record which model powers your Companion and the reasoning behind it. Complete it in one go and save as a md file. Save the md file within the same source folder you've given Evidence Companion access.
💬 Reflection prompt: Take one genuine rough note from your work and run it through two different model families — for example your Companion's model and one other you can access. Read both drafts side by side. Which "house voice" fits the way you want your evidence to read? Because raw capability has largely converged, this tone test is often the decision that actually matters.
⏭️ Up next - Lesson 4: You now understand the model layer — the LLM that reasons inside your automation — and you have chosen the one that powers your Companion. The next lesson stays in the model layer but widens it: models that go beyond text, working with images, documents, voice, and video, and what your Companion could ingest as a result.