Foundation & Frontier Models

Frontier models are the most capable general-purpose AI models available at a given moment: the small set trained with the largest compute budgets, which define what the best systems can do. Every frontier model is a foundation model, meaning one large model pretrained on broad data and then adapted to many tasks. Most foundation models are not frontier. They are smaller, older or open-weight models that cost less to run and do most production work well.

The word "frontier" names a position, not a product. A model at the frontier this year usually sits in the middle tier a year later, and the gap between the leaders and the rest is measured in months. That one fact should shape how you build on any of them.

What makes a model a foundation model#

The term comes from Stanford's 2021 report "On the Opportunities and Risks of Foundation Models", written by Bommasani and 113 co-authors, which described models "trained on broad data at scale" that "are adaptable to a wide range of downstream tasks". Before that, a company trained one model for translation, another for summarization and another for sentiment. Now one pretrained base can do all three.

Pretraining is self-supervised: the model reads a vast amount of text and code and learns by predicting the next token, so the data labels itself and no person has to. The cost is concentrated in that one run. Epoch AI estimated in June 2024 that the cost of training frontier models had been growing about 2.4x a year, and projected that the largest runs would cost over a billion dollars by 2027. Adapting the result is small by comparison, which is why a handful of labs train bases and everyone else builds on them. Build on a base unless pretraining is your product. Every system in our case studies is built on an existing base model.

Adaptation comes in three forms of rising commitment. Prompting, including examples in the context through in-context learning, changes only the input. RAG leaves the model alone and supplies your documents at query time. Fine-tuning moves the weights to fix a format, register or narrow behaviour.

Who decides which models count as frontier models#

There is no official list. Anthropic, Google, Microsoft and OpenAI founded the Frontier Model Forum in July 2023 as an industry body for frontier AI safety, and its site now lists Amazon and Meta as members too, but it does not publish a definition that sorts models in or out.

Regulators use training compute as a proxy. Under the EU AI Act, a general-purpose model trained with more than 10^25 floating-point operations is presumed to carry systemic risk, and the obligations for general-purpose model providers have applied since 2 August 2025. The Commission says that threshold is under review. Treat the frontier as whichever few models currently lead on the tasks you care about, measured on your own evals, and expect that set to change within the life of your project.

Closed API or open weights#

Foundation models ship two ways. Closed models are reachable only through the vendor's API. You get the strongest capability with nothing to operate, and in exchange your data passes through a third party, the price and behaviour can change under you, and the model is retired on the vendor's schedule. Open-weight models publish their weights. You can run them in your own cloud or on your own hardware, fine-tune them freely, pin a version for as long as you like, and keep regulated data inside your perimeter. You also own the GPUs, the serving stack and the patching.

The capability gap has narrowed. Epoch AI's November 2024 report put the best open models 5 to 22 months behind the best closed ones, about a year on average. Its May 2026 update, on its own capabilities index, found open-weight models lagging frontier closed models by an average of about four months since January 2026.

For some systems the choice is made by where the data may go, not by capability. On the on-device meeting transcription we built with a SalesTech client's team, cloud transcription was a disqualification for customers with data-residency obligations, so a compact speech model runs on the laptop and no audio is uploaded. There is no upload path, so keeping audio on the device does not depend on anyone following a policy.

Build so the model can change#

Each of our designs puts correctness outside the model, so the model can change without changing what the system guarantees. On the DevOps agent, the agent is treated as untrusted and every step is either checkable by a machine or reviewed by a person, so nothing depends on the model being correct and an engineer does not need an opinion of the model to decide whether to run it. On the sales roleplay coach, the sales coach we built, the model writes the feedback but a versioned rubric sets the score, because if the model did both, a model upgrade would shift every historical comparison.

Bigcircle recommends three habits that follow from that. Prototype on a frontier model first, because if the best available model cannot do the task, no smaller one will, and you have removed capability as the open question. Once the workflow works, write evals and use them to move the easy traffic to smaller or open models, with model routing if the split is per request, because the evals are what tell you which requests can move. And keep the model name out of your business logic, because a model swap should be a configuration change plus an eval run, not a code change.