Lesson 5 of 57
Frontier Models
News about AI uses the phrase "frontier model" all the time, especially when the subject is large language models. It sounds like an official category. It is not one.
This lesson explains what the term means, how a model comes to be called a frontier model, and why a model stops being one. After this lesson you can read a headline such as "a new frontier model was released" and know what is being claimed.
| Idea | Short meaning |
|---|---|
| Frontier model | One of the most capable general-purpose models available right now |
| General purpose | One model for many kinds of task, not a specialist |
| Benchmark | A fixed set of tasks given to every model and scored the same way |
| AI lab | A company or group that trains models |
1. What "frontier" means
A frontier is a leading edge. A frontier model is one of the best models available at the moment.
Picture every model that anyone can use, placed on a line from less capable on the left to more capable on the right. Most models sit somewhere in the middle. A few sit at the far right. Those few are the frontier.
flowchart LR
A["Older and smaller models"] --> B["Capable models"] --> C["The frontier<br/>the most capable models<br/>right now"]Two properties go with the term:
- Leading edge. A frontier model is among the most capable models that exist right now.
- General purpose. It is not a specialist. It is one model that is used for whatever the task is: writing, coding, analysis, and more.
Frontier models are usually very large. They have very high numbers of parameters, so they run in data centres and not on your own computer.
2. How a model earns the name: two signals that must agree
Nobody measures "best" with one number. Two different signals are used, and both have to point the same way.
| Signal | What it is | What it captures |
|---|---|---|
| Benchmarks | The same tasks, given to every model and scored the same way | Results that can be compared in a table, often called a leaderboard |
| Daily use | What people who work with the model every day say about it | The part that no score sheet captures |
Daily-use reports sound like this: "I switched over last week. It follows what I meant." or "Same for me on long documents."
A model is called a frontier model when both signals agree:
| Check | Result |
|---|---|
| Is it at the top of the benchmark leaderboards? | Yes |
| Do people who use it prefer it? | Yes |
| Frontier model | Agreed |
A model that scores well on benchmarks but disappoints in daily use does not earn the name. Neither does a model that people like but that falls behind on every benchmark.
3. It is not a fixed or official term
Nobody publishes the list of frontier models. There is no committee and no certificate. The category is agreed on by the people in the field. It is not awarded.
This means the term is somewhat loose. Different people may draw the line in slightly different places.
4. Membership keeps changing
The label is temporary. New models arrive and older ones leave the category.
flowchart LR
N["Arriving<br/>a new release,<br/>and then the next one"] --> F["The frontier category"]
F --> L["Leaving<br/>still a good model,<br/>no longer at the edge"]When a new, better model is released, it joins the frontier. An older model that was at the edge last year is still a good model. It is no longer the most capable, so people stop calling it a frontier model.
Membership changes without anyone announcing it. Labs release new versions every few months, so today's frontier model can be an ordinary model within a year.
5. Who people mean when they say it
When someone says "frontier models" today, they usually mean the latest models from a small group of AI labs.
| Lab | Their frontier models |
|---|---|
| OpenAI | The latest GPT models, which run ChatGPT |
| Anthropic | The latest Claude models |
| The latest Gemini models | |
| A few more | Other labs, including several from China such as DeepSeek, Alibaba (Qwen), and Moonshot AI (Kimi) |
Only a small number of organisations can build frontier models, because training one takes very large amounts of money, computing hardware, and data.
Several of the Chinese labs release their models openly, so anyone can download and run them. The models from OpenAI, Anthropic, and Google mentioned above are reached through those companies' apps and services.
Which model counts as the best changes often. Treat any list like the one above as a snapshot.
Interview points
- A frontier model is one of the most capable general-purpose models available at a given time.
- The term has two parts: leading edge in capability, and general purpose instead of specialist.
- Two signals decide it: benchmark results and the preference of people who use the model daily. Both have to agree.
- A benchmark gives the same tasks to every model and scores them the same way, so results can be compared.
- No organisation publishes an official list. The category is agreed on informally.
- Membership changes as new models are released. An older model remains useful but is no longer called frontier.
- The term usually refers to the latest models from OpenAI, Anthropic, Google, and a few other labs, several of them Chinese.
- Building a frontier model needs very large budgets, hardware, and data, which is why few labs do it.
Summary
- "Frontier" means the leading edge of what models can do right now.
- Frontier models are general purpose and usually very large.
- Benchmarks and daily use together decide which models count.
- The label is informal and temporary.
- The usual names are the latest GPT, Claude, and Gemini models, plus models from a few other labs.