What is frontier AI? Definition and EU access

Frontier AI: the definition

"Frontier AI" is the name for the most capable general-purpose AI models of the moment. The definition most often cited comes from the UK government's October 2023 discussion paper, written ahead of the Bletchley Park AI Safety Summit: frontier AI means "highly capable general-purpose AI models that can perform a wide variety of tasks and match or exceed the capabilities present in today's most advanced models" (GOV.UK).

Three things follow from that sentence:

• It is general-purpose. The paper explicitly leaves out powerful narrow systems such as AlphaGo, AlphaFold or DALL·E 3, because they cannot perform as wide a range of tasks. A frontier model writes, codes, reasons, summarises and calls tools.
• It is a moving target. "Today's most advanced models" changes every few months. A model that was frontier in 2023 is not frontier in 2026; the label describes a position, not a product.
• It is not tied to a technology. In 2023 it meant large language models, the ones behind ChatGPT, Claude and Bard. The paper notes that a future frontier system might rest on something else.

The term was popularised a few months earlier. A July 2023 paper, "Frontier AI Regulation: Managing Emerging Risks to Public Safety", co-written by researchers from AI labs and universities, used it for highly capable foundation models that could pose severe risks (arXiv). On 26 July 2023, Anthropic, Google, Microsoft and OpenAI launched the Frontier Model Forum, describing frontier models as large-scale machine-learning models that exceed the capabilities of the most advanced existing models and can perform a wide variety of tasks (announcement).

How regulators draw the line

A capability definition is hard to enforce, so lawmakers use a proxy: the amount of computation spent training the model, counted in floating-point operations (FLOP).

The EU threshold is a presumption, not a hard line. The Commission can designate a model below it on the basis of its capabilities and reach, a provider above it can argue its model carries no systemic risk, and the figure itself can be amended by delegated act (AI Act Service Desk). The obligations for general-purpose AI model providers have applied since 2 August 2025; the Commission's enforcement powers start on 2 August 2026, and models already on the market before August 2025 have until 2 August 2027 (European Commission). In California, SB 53 targets the same small group of developers through transparency rather than pre-approval (Governor of California).

Either way, "frontier" in law means a handful of models from a handful of companies. What those rules mean for a company that only calls a model through an API is covered on our AI Act page: the general-purpose AI chapter puts its obligations on the model's provider, while yours depend on your use case.

Who builds frontier AI in 2026

Frontier models come from two places, and they ship in two different ways.

• Closed models from US labs. OpenAI, Anthropic and Google DeepMind sell access to their best models only through their APIs and those of their cloud partners. You never get the weights; you get an endpoint run by a US company, even when it sits in a European region.
• Open-weight models, mostly from Chinese labs. DeepSeek, Qwen (Alibaba), Z.ai (GLM) and Moonshot AI (Kimi) publish the weights of very large models under licences that usually allow commercial use. Their own APIs run outside Europe, but because the weights are public, anyone with the hardware can serve them.

Europe has its own labs. Mistral AI, in Paris, publishes open-weight models and sells larger ones through its API; the Swiss AI Initiative (EPFL, ETH Zurich, CSCS) released Apertus, a fully open model whose training data is documented. Neither is a substitute for every closed US model, but together with the open-weight releases they make a European route to very capable AI realistic.

Here are some of the largest open models in our catalogue, with where each one runs when you call it through Frontière AI:

Parameter counts are total parameters, read from each model's page; size is not a ranking, and we do not claim any of these beats a given closed model. Mistral Large 4 is still labeled "Public Preview" by Mistral: under § 4.3 of Mistral's terms, data sent to preview models may be used for training, and opt-outs do not apply — see its model page. Rows labeled fast access run at US companies and are never presented as sovereign.

Why open weights change the sovereignty question

With a closed frontier model, the question "who processes my data?" has one answer: the lab, or a cloud provider it has chosen. You can pick a European region, but not a European company, and US law follows the company: the CLOUD Act lets US authorities require a provider under US jurisdiction to hand over data it controls, wherever the servers are (our GDPR vs. CLOUD Act guide goes through the detail).

Open weights break that link. The model is a file; whoever runs it decides where it runs. A DeepSeek or GLM model served by a French or German company, in its own data centres, is the same model as the one on the lab's API, with a different operator and a different jurisdiction. That is what makes "EU-hosted AI models" a real category rather than a marketing line. Our guide to open-weight AI explains what the weights are and what the licences allow.

What "EU-hosted AI models" should mean

An EU-hosted AI model is a model whose inference — the computation that turns your prompt into an answer — runs in data centres located in the European Union or a country with an EU adequacy decision. That is the minimum. For the label to protect anything, check three things separately:

1. The location. Which data centres actually run the request, and whether the provider commits to them. Some publish a city; others route "dynamically" and warn that the region can change.
2. The company. Who operates the servers and who controls that company. A European region inside a US company answers where, not who.
3. Your prompts. Whether inputs and outputs are kept, for how long, and whether they can be used for training.

A provider can pass one test and fail another, so "hosted in Europe" on a product page is the start of the check, not the end. We compared twelve providers on these three points in European inference providers in 2026, each fact quoted from the provider's own pages; for the same question organised by model, see EU-hosted AI models, family by family.

What an EU LLM gateway is

An EU LLM gateway — also called a European AI gateway — is a single API placed in front of several European inference providers. You keep one key and one SDK, and the gateway sends each request to a provider that meets the jurisdiction you asked for. Compared with signing up to each provider, it gives you:

• One integration. An OpenAI-compatible endpoint, so existing code changes its base URL and key, nothing else. Our documentation shows the two lines.
• A choice of models across providers. No single European provider serves every model; a gateway lets you call Mistral, Qwen, GLM, DeepSeek or Llama from one account.
• Failover that respects jurisdiction. When a provider is down, the request moves to the next eligible one. The point of a European gateway is that "next eligible" never silently means a US provider.
• Evidence on every call. The provider that served the request, and whether it is sovereign, should come back with the answer, not only in a policy document.

The alternative is a gateway you run yourself, such as LiteLLM: more control, more operations, and the providers' contracts remain yours to sign. We compare both paths in open-source AI gateways and in our LiteLLM guide, which uses Frontière AI as an upstream. Note that a general-purpose router with a European region option is not the same thing: if the default route can leave Europe, sovereignty depends on every request being configured correctly. Our OpenRouter alternatives article looks at which routers keep traffic in Europe by default, and our guide to choosing an EU LLM gateway compares the European ones point by point.

Sovereign AI, in three levels

"Sovereign AI" is used for three different things, and it helps to say which one you mean:

1. Sovereign infrastructure. A country or a company owns the GPUs, the data centre and the operations. It is the strongest form of control and by far the most expensive; our analysis of building your own cluster versus using an API puts numbers on it.
2. Sovereign models. Models trained in Europe, by European organisations, on documented data — Mistral's models or Apertus, for instance. This answers who built the model, not who runs it.
3. Sovereign access. Using capable models, wherever they were trained, through European companies running them in Europe, with no prompt retention. It is the level most teams can reach this week.

Frontier AI and sovereign AI meet at the third level: since frontier-scale open-weight models exist, you do not have to choose between capability and jurisdiction for most workloads. What sovereign access does not give you is immunity. A European provider sits outside the CLOUD Act's direct reach, not outside every foreign law — OVHcloud and Scaleway, for example, each have a US entity. Our sovereignty methodology explains how we score each model against these questions, and what sovereign AI means goes deeper into the term, Europe's public programmes and the certifications.

Frontier AI under European jurisdiction: where Frontière AI fits

The name is a pun that is also the brief. "Frontière" is French for border: frontier models, served on this side of it. Frontière AI is a European LLM gateway run by FLEECE AI SASU, a French company. It does not own GPUs; it routes to European providers and labels every model by jurisdiction.

• 17 EU-sovereign models live today, served by Scaleway (France), Mistral AI (France), Infomaniak (Switzerland) and AKI.IO (Germany). A model labeled sovereign stays on sovereign providers by default, failover included; if none is available, the call fails rather than leaving Europe.
• Fast-access models are labeled. Some new releases are only available from US companies at first. We offer them under a separate "fast access" label, never as sovereign, and an API key can be restricted to sovereign models only.
• Proof on every response. Each answer carries X-Routed-Provider and X-Sovereign headers, so your logs show where each request ran.
• No prompt storage. Prompts and completions are not written to our database. The gateway itself runs on Vercel's Paris region — a US company, in transit only — as our privacy policy states.
• Prepaid, no subscription. Top up from €10, VAT included, and pay per token; open to companies and individuals.

Buying directly from a single provider is cheaper, and we say so on our comparison pages. A gateway earns its margin when you want several providers behind one key, with the jurisdiction named on every call. Browse the model catalogue or read why teams use Frontière AI.

FAQ