AI Glossary / AI foundations / Beginner

Frontier model

Definition

A frontier model is an AI model at or near the leading edge of capability at a given time, usually a large general-purpose model trained with very large amounts of data and computing power. It is not a technical standard or a fixed list, and no single body defines it. Regulators, labs and analysts use the label differently, and a model that counts as frontier today may be ordinary within a few years.

What it means

For a business reader, a frontier model is the strongest class of general-purpose AI model on the market right now. These are the models that set the pace on writing, coding, analysis, reasoning over long documents, and increasingly on images, audio and tool use. They are usually built by a small number of well-funded organizations because training them is very expensive.

The word is relative. It describes a moving position, not a permanent category. Capabilities that were frontier a few years ago are now available in small, cheap models. When a vendor says a product runs on a frontier model, ask which model, what it was tested on, and whether the claim still holds for the version you will actually use.

How it works

Frontier models are usually large neural networks, most often built on the transformer design, trained on very large collections of text, code and other data. Training adjusts billions of internal numbers, called parameters, so the model gets better at predicting what comes next. A second stage, often using human feedback, shapes the model to follow instructions and avoid harmful output.

What pushes a model to the frontier is scale and refinement: more data, more computing power, better training methods and more careful tuning. These ingredients tend to improve performance on many tasks at once, which is why a frontier model can draft a contract clause, explain a spreadsheet and write code with the same weights. It can also show abilities its builders did not plan for, which is one reason testing matters.

You rarely run a frontier model yourself. You reach it through a vendor's interface or API, and the vendor controls updates, pricing and safety measures. Policy discussions sometimes use rough thresholds, such as the amount of computing used in training, to decide which models get extra oversight. These thresholds vary by jurisdiction and are revised over time.

Three practical examples

• Finance team: An accounts team uses a frontier model to read long supplier contracts and flag unusual payment terms. A reviewer checks every flag against the source text before anything is escalated, because the model can misread clauses.
• Retailer: A retailer routes its hardest customer questions, such as multi-step returns with partial refunds, to a frontier model. Simple order-status questions go to a smaller, cheaper model, since they do not need frontier ability.
• Software team: A small development team uses a frontier model to draft code and write tests. Engineers review and run everything, and they keep a fallback model in case the vendor retires or changes the one they rely on.

Think of it this way

Think of a frontier model as the fastest car in a racing season: it holds the record for now, costs the most to build, and will probably be matched by ordinary road cars in a few years. The analogy stops working because a car's speed is one number, while a model's strength varies by task, so the best model for one job may not be best for another.

Worked example

A support desk handles 10,000 customer messages a month. A frontier model costs 0.02 dollars per message and resolves 90 percent correctly. A smaller model costs 0.002 dollars per message and resolves 82 percent correctly. Each wrong answer costs 3 dollars in staff time to fix.

Frontier model: 10,000 x 0.02 = 200 dollars in model fees. Errors: 10,000 x 10 percent = 1,000 wrong answers. 1,000 x 3 = 3,000 dollars in fixing. Total: 3,200 dollars.

Smaller model: 10,000 x 0.002 = 20 dollars in model fees. Errors: 10,000 x 18 percent = 1,800 wrong answers. 1,800 x 3 = 5,400 dollars in fixing. Total: 5,420 dollars.

The frontier model is cheaper overall by 2,220 dollars a month, even though its fee is ten times higher. If the smaller model resolved 89 percent instead, its errors would cost 3,300 dollars plus 20 in fees, 3,320 dollars total, and the frontier model would lose its edge. The decision depends on the measured error rate and the cost of an error, not on the label frontier.

Illustrative case study

Fictional case: Larkspur Mutual, a made-up regional insurer, adopted a frontier model to summarize claim files. Early results looked excellent, and the team planned to use it for all claim types. In month two the vendor released an updated version of the model. Summaries of water-damage claims began leaving out policy exclusions, a detail the old version had usually included. A routine sample review caught the change after a week, and no payments had been decided on the faulty summaries. Larkspur fixed it by pinning the model to a named version where the vendor allowed it, building a test set of 200 past claims with known correct summaries, and re-running that set before accepting any update. It also required an adjuster to confirm exclusions on every summary. Later updates were tested first and adopted only when scores held.

Common misconceptions and mistakes

A frontier model is always the right choice for a business task. Many tasks are handled well by smaller, cheaper models, and the best choice depends on measured accuracy, cost, speed and risk for your specific use.

Frontier means the model is correct and trustworthy. Frontier models still make confident mistakes, so outputs that matter need checking against sources by a person.

Frontier is a fixed, official category. No single authority defines it, and the line moves as capabilities improve and as different regulators and vendors set their own criteria.

A model that is frontier today will stay ahead. Leadership changes often, and today's frontier abilities usually reach smaller models within a few years.

Business use

Start by writing down the task, a set of real examples with known good answers, and a success measure. Test one frontier model and one smaller model on that set, compare accuracy, cost per task and speed, and pick the cheapest option that meets your quality bar. Keep a human check on any output that affects customers, money or compliance.

Track error rate, cost per completed task, and how often reviewers correct the output. Costs include usage fees, integration work and review time. Risks include vendor changes to a model without notice, dependence on one supplier, data privacy when sending information to a vendor, and regulatory duties that may attach to certain high-capability models. Review your setup whenever the vendor updates the model.

People also ask

Is a frontier model the same as a large language model?

No. Many frontier models are large language models, but the term describes position at the leading edge of capability, not a design. Some frontier models also handle images, audio and video.

Do I need a frontier model for my business?

Often not. Many routine tasks, such as classifying emails or extracting fields from forms, work well with smaller models. Use a frontier model when testing shows the task needs the extra capability.

Why do regulators care about frontier models?

Their broad capabilities and wide use can create risks that are hard to predict. Some rules therefore add testing, documentation or reporting duties for the most capable general-purpose models, and the details differ by jurisdiction.

Check your understanding

In the worked example, what is the total monthly cost of the frontier model?

Answer: 3,200 dollars: 200 dollars in fees plus 3,000 dollars to fix 1,000 wrong answers at 3 dollars each.

Why did the smaller model cost more overall in the worked example?

Answer: Its lower accuracy produced 1,800 errors instead of 1,000, and the extra fixing cost outweighed the fee savings.

Why is frontier called a relative term?

Answer: It describes being at the leading edge at a given time. As capabilities spread, yesterday's frontier model becomes ordinary.

What did Larkspur Mutual change after the model update caused missing exclusions?

Answer: It pinned the model version where possible, built a 200-claim test set to re-run before accepting updates, and required adjusters to confirm exclusions.

Related terms

• Foundation model
• Large language model
• Generative AI
• Transformer
• Training data
• Parameters
• Scaling laws
• Benchmark

Learn next

Build a small test set of 20 to 50 real tasks from your own work and try two models on it, noting accuracy and cost. Then read the entries on foundation models and large language models for the wider picture, and on benchmarks to learn how model claims are tested.

More in AI foundations

• Feature attribution - Feature attribution is a family of explainability methods that assign each input feature a score showing how much it pushed a model's output up or down for...
• Fine-tuning - Fine-tuning is additional training that adapts a pretrained model to a particular task, domain, or desired behavior.
• Floating-point operations (FLOPs) - Floating-point operations (FLOPs) are individual arithmetic steps, such as adding or multiplying two decimal numbers, that a computer performs.
• General-purpose AI model - A general-purpose AI model is an AI model trained on very large and varied data so that it can perform many different tasks, rather than one narrow job.
• Generalization - Generalization is a model's ability to perform well on new cases it has never seen, rather than merely repeating answers about the examples it was trained on.
• Generative AI - Generative AI is artificial intelligence that produces content, such as text, pictures, audio, video or computer code, using patterns learned from data.

Sources

Stanford HAI, AI Index: https://hai.stanford.edu/ai-index NIST, AI Risk Management Framework: https://www.nist.gov/itl/ai-risk-management-framework European Union, Artificial Intelligence Act (Regulation (EU) 2024/1689): https://eur-lex.europa.eu/eli/reg/2024/1689/oj The sources were consulted for background on how capable general-purpose models are tracked, how risk is managed and how regulators treat high-capability models; the examples, analogy, worked example and case are original educational illustrations. Checked 5 Nov 2026.

Previous: Frontier Next: Full functional dependency