Research · Generative & Agentic AI

Foundation Models & LLMs

Foundation models are large AI models trained on broad data that can be adapted to many tasks. Large language models (LLMs) are the best-known type. Choosing and managing them well is now a core technology decision.

The Model Portfolio Grid

Most enterprises no longer pick one model. They run a portfolio that mixes hosted frontier models, smaller fast models, reasoning models that spend extra compute working through a problem before they answer, and open-weight models they can run on their own infrastructure. The Model Portfolio Grid, a CorpExcellence framework, places each use case on two questions. How much reasoning does the task demand? How sensitive is the data, and how much control do you need over where the model runs? The answers point to a tier, and each tier points to a default model class and deployment option.

The four tiers

• Commodity tier. Routine work on low-risk data, such as tagging tickets or pulling fields from forms. Use the cheapest fast model that passes your evaluations. Paying for frontier capability here is a common source of waste.
• Frontier API. Complex analysis or multi-step planning on public or low-sensitivity data. Hosted frontier and reasoning models earn their price here. Many reasoning models let you set how much effort they spend, so match effort to the task instead of defaulting to the maximum.
• Private efficient. Routine work on sensitive data, such as internal case notes or regulated records. Smaller or open-weight models hosted in your own cloud tenancy or data center keep data inside your boundary at a predictable cost.
• Controlled frontier. Hard problems on sensitive data. Options include frontier models under private deployment terms that exclude training on your data, or large open-weight models you host and patch yourself. Expect the highest cost and the most review in this tier.

Making the grid operational

The grid works only when shared services sit behind it. An AI gateway routes each request to the approved model for its tier and records cost per use case. A standing evaluation set built from your own tasks decides which models qualify for each tier, because public benchmarks rarely match enterprise work. A model register tracks license terms and deprecation dates, since providers retire models on their own schedule and a replacement can behave differently. Re-run the evaluations whenever a model in the portfolio changes.

Which requests actually need a premium model?

AI Inference Cost Control covers model right-sizing, routing patterns and cascades, with a model eligibility decision matrix and a routing policy template.

Get the full report →

Where the industry is now

Two shifts shape portfolio decisions as of September 2026. First, open-weight models from major labs are credible options for the private tiers. OpenAI, for example, released its gpt-oss open-weight reasoning models under the Apache 2.0 license in August 2025. Second, regulation now reaches model providers directly. Under the EU AI Act, obligations for providers of general-purpose AI models have applied since August 2, 2025, and the European Commission can enforce them, including through fines, from August 2, 2026 (Commission guidance). Enterprises that fine-tune or substantially modify a model should check whether they take on provider duties.

Common pitfalls

• Choosing a model from leaderboard rankings instead of your own evaluation set.
• Hard-coding one provider’s API so that switching requires rewriting applications.
• Treating “open-weight” as “open source” without reading the license.
• Skipping re-evaluation when a provider updates a model behind the same name.