Frontier model

What is a frontier model?

A frontier model is one of the handful of AI models that, right now, can do more than anything else on the market. The word frontier describes a position, not a property: it says where a model sits relative to the others at a given moment, and that position does not last.

The term came into common use in 2023, when Anthropic, Google, Microsoft and OpenAI set up the Frontier Model Forum and described frontier models as large-scale models that exceed the capabilities of the most advanced existing models and can perform a wide variety of tasks. Their own definition adds that this is a group that will shift over time as the field progresses. A frontier model is a general-purpose model, so it is also a foundation model and, in AI Act terms, a GPAI model. What sets it apart is only that it is at the top of the pile today.

In practice the frontier is a short list. In 2026 it means the top models of a few American labs, with Chinese models close enough that the Stanford AI Index measured the lead at under three percent in March 2026. Almost all of them are closed: you reach them through an API or a hosted product, and you cannot download the weights.

Who builds frontier models, and what it costs

Training a frontier model is expensive enough that only a few organisations do it. Epoch AI, a research group that tracks compute and cost trends in AI, estimated in 2024 that the hardware and energy cost of the final training run of frontier models had grown by roughly 2.4 times per year since 2016, and projected that the largest runs would pass one billion dollars by 2027. The European Commission, explaining the AI Act's threshold for the largest models, put the cost of training at that level at tens of millions of euro and noted that such models come from a handful of companies. The AI Index 2026 counts over ninety percent of the notable models released in 2025 as coming from industry, with universities and public labs sharing the rest.

That is why frontier is also shorthand for a small group of labs. The Frontier Model Forum now counts Amazon, Anthropic, Google, Meta, Microsoft and OpenAI as members, and its membership criteria ask for the resources to develop or deploy frontier models at scale.

Why the label moves

Last year's frontier is this year's mid tier. Each lab releases its top model, then a cheaper and faster model a few months later that matches most of what the top model could do, sometimes through distillation of the larger one. The model you paid top rates for in 2025 is now the default or budget option in the same vendor's price list.

Open-weight models follow the same curve with a delay. Epoch AI measured in May 2026 that since January the best open-weight models had lagged the best closed models by about four months on its aggregate capability index, slightly more than the roughly three months it had measured in late 2025. The strongest open-weight models at that point came from Chinese labs such as Moonshot, Zhipu and Alibaba. So whatever the frontier can do today, you can expect to run something close to it on your own hardware within half a year.

Frontier model versus open-weight model

The two categories pull in opposite directions on one dimension: control versus capability.

Capability. A frontier model is, by definition, the most capable thing you can call. On hard judgement (a messy contract, an ambiguous complaint, a multi-step analysis where one wrong step ruins the result) the difference against a model a few months behind shows up in the error rate. On routine work it usually does not.

Control. An open-weight model runs where you put it. Your data stays inside, you pin the exact version, and the vendor cannot retire it, raise the price or change its behaviour underneath you. With a frontier model you accept all of that in exchange for the extra capability. Frontier models are almost always closed. When a big lab does release open weights, it is usually a smaller model than its best one.

There is a transparency cost too. The AI Index 2026 reports that the Foundation Model Transparency Index fell to 40 points from 58 the year before, and that the most capable models often disclose the least. Ask for the model card and the documentation the AI Act requires, and do not assume the biggest model comes with the best paperwork.

Frontier models in regulation and safety frameworks

The AI Act does not use the word frontier, but its tier of general-purpose AI models with systemic risk is aimed at the same group. Article 51 presumes a model has high-impact capabilities when its training used more than 10^25 floating-point operations, and the Commission can raise or lower that threshold as the technology moves. A provider that crosses it must notify the Commission within two weeks (Article 52), and Article 55 then adds the duties of the systemic-risk tier: evaluation with adversarial testing, risk mitigation, incident reporting to the AI Office, and cybersecurity for the model and its infrastructure. Annex XIII adds a reach criterion: a model offered to at least 10,000 registered business users in the EU is presumed to have a high impact on the internal market.

The labs publish their own frontier safety frameworks alongside this. Anthropic's Responsible Scaling Policy, first published in September 2023 and at version 3.4 since July 2026, defines capability thresholds (in areas such as biological weapons or autonomous AI research) that trigger stricter security and deployment standards once a model crosses them. Google DeepMind's Frontier Safety Framework, introduced in May 2024 and updated in early 2025, works the same way with what it calls critical capability levels. Both are voluntary, and the EU's Code of Practice for general-purpose AI, published in July 2025, describes practices of the same kind for providers of the largest models.

None of these obligations fall on you as a user. They are the questions to ask a provider, and the reason the largest models come with evaluation reports at all.

When do you use a frontier model?

Rarely for routine work. Classifying incoming mail, extracting fields from invoices, drafting a standard reply, summarising a meeting: a model one or two tiers down does these about as well, at a fraction of the price, and can often run in your own environment.

You do want the frontier for the hard parts: the ten percent of cases where judgement matters, where a mistake is expensive, or where the task chains many steps together. Reviewing a non-standard contract clause, deciding whether a complaint needs escalation, writing code that touches your ERP, checking another model's output before it goes to a customer.

A common pattern is therefore frontier for the hard parts, open weights or a small model for the rest. Take 10,000 supplier emails a month, each about 1,000 tokens in and 200 tokens out. On one vendor's current price list, the top tier costs roughly 10 dollars per million input tokens and 50 per million output tokens; the smallest tier costs about a tenth of that. Everything through the frontier model comes to about 200 dollars a month, everything through the small model to about 20. Routing the 90 percent that are routine to the small model and escalating the rest comes to just under 40 dollars, with the hard cases still handled by the best model you can get. At this volume the money is small either way. At a hundred times the volume, or with an open-weight model on hardware you already own for the routine share, the split is what makes the project affordable.

The router that decides which case is hard can be a rule, a small classifier or the small model itself. If it errs, let it err in the safe direction: when in doubt, escalate.

What to watch out for with frontier models

The model under the name changes. Vendors retire versions and ship new ones on their own schedule. Pin an exact version where you can, keep an evaluation set of your own cases, and rerun it before every switch.

Price and behaviour are the vendor's to change. A frontier model is a subscription to someone else's roadmap. Budget for that, and keep the routine share of your workload on something you can replace within a week.

Frontier is not a requirement. A supplier who insists that only the newest top model can do your job is either selling you something or has not tested the cheaper tier. Ask for the comparison on your own data.

Your data goes to the lab. Frontier models are hosted. Check the data processing terms, the region and the retention period before personal or confidential data flows through, the same as with any other external processor.