# Frontier Models Source: John Bradshaw — https://bradshaw.cloud URL: https://bradshaw.cloud/strategy/frontier-models/ Type: insight Date: 2026-04-21 Updated: 2026-08-15 Tags: Models, AI > The most capable AI models available — and when a frontier model earns its cost vs when a small specialised model wins. ## Frontier models explained ### Frontier models: the F1 cars of AI Frontier models are the most capable general-purpose AI models available — Claude Opus, GPT-5, Gemini Pro. They're trained on trillions of tokens with billions of parameters at a cost of $100M+. They can reason, write, code, analyse images, and handle tasks nobody explicitly trained them to do. **Business take:** The business analogy: an F1 car. Astonishing engineering, extraordinary performance, astronomical running costs. Perfect for the qualifying lap. Terrible for the school run. ### When frontier models earn their cost Frontier models are the right answer when the task requires broad reasoning, handles novel situations, or resists being reduced to a well-defined pattern. They're also the default for low-volume, high-value internal use. **Business take:** Use a frontier model when the cost of a wrong answer exceeds the cost of the API call. Legal analysis, strategic decision support, complex customer escalations, creative work. Use a small model when the task is well-defined and repeatable. ### Frontier models in 2026: the key players The proprietary frontier is led by Anthropic (Claude Opus), OpenAI (GPT-5), and Google (Gemini Pro). The open-weight frontier — models you can self-host — is led by Meta (Llama 4), Mistral, Alibaba (Qwen), and DeepSeek. **Business take:** The strategic question: proprietary frontier models give maximum capability but lock you into a vendor API. Open-weight frontier models sacrifice some capability but give you control over deployment, cost, and data residency. ### The real architecture: intelligent model routing Production systems don't pick one model. They route each request to the cheapest model that can handle it. Simple classification goes to a 7B model ($0.001). Complex reasoning goes to a frontier model ($0.05). The router decides in milliseconds. **Business take:** This is how you get frontier capability without frontier costs. The router is the economic brain of the system — it turns a fixed "which model?" decision into a per-request optimisation problem. ## Frequently asked questions ### What is a frontier model? A frontier model is one of the most capable general-purpose AI models available at any given moment — trained on trillions of tokens, with billions of parameters, at a cost north of $100M. They reason, write, code, analyse images, and handle tasks nobody explicitly trained them to do. The useful analogy is an F1 car: astonishing engineering, extraordinary performance, astronomical running costs. Perfect for the qualifying lap, terrible for the school run. ### When does a frontier model earn its cost? When the cost of a wrong answer exceeds the cost of the API call. That covers broad reasoning, novel situations, and anything that resists being reduced to a well-defined pattern — legal analysis, strategic decision support, complex customer escalations, creative work. It also covers low-volume, high-value internal use, where the token spend is trivial next to the value of the output. Where the task is well-defined and repeatable, a small model wins. ### What is the difference between proprietary and open-weight frontier models? Proprietary frontier models give you maximum capability behind a vendor API. Open-weight models — the ones you can download and self-host — give up some capability in exchange for control over deployment, cost, and data residency. The strategic question is not which scores higher on a benchmark; it is whether the capability gap is worth the lock-in, and whether your data is allowed to leave your estate at all. ### What is model routing, and why do production systems use it? Production systems do not pick one model — they route each request to the cheapest model that can handle it. Simple classification goes to a small model at a fraction of a penny; complex reasoning goes to a frontier model at fifty times the price. The router decides in milliseconds. This is how you get frontier capability without frontier costs: it turns a fixed "which model?" decision into a per-request optimisation. ### Does every AI task need a frontier model? No, and defaulting to one is the most expensive habit in enterprise AI. Using a frontier model to classify sentiment or extract fields from a form is couriering a Post-it note. Most production traffic is well-defined and repeatable, which is exactly what small specialised models are good at. Reserve the frontier tier for the fraction of requests that genuinely need it — and route the rest. ## Diagrams Standalone SVG diagrams, licensed CC BY 4.0 with attribution to John Bradshaw — link back to https://bradshaw.cloud/strategy/frontier-models/ when embedding. - [Frontier models: the F1 cars of AI (SVG)](https://bradshaw.cloud/strategy/frontier-models/diagrams/what.svg) - [When frontier models earn their cost (SVG)](https://bradshaw.cloud/strategy/frontier-models/diagrams/when.svg) - [Frontier models in 2026: the key players (SVG)](https://bradshaw.cloud/strategy/frontier-models/diagrams/landscape.svg) - [The real architecture: intelligent model routing (SVG)](https://bradshaw.cloud/strategy/frontier-models/diagrams/routing.svg) ## Related - [AI fundamentals](https://bradshaw.cloud/strategy/ai-fundamentals/) — the building blocks that sit under every model choice. - [AI cost curves](https://bradshaw.cloud/strategy/ai-cost-curves/) — the economics that decide frontier vs self-hosted. - [Agentic AI patterns](https://bradshaw.cloud/strategy/agentic-ai/) — the orchestration that routes each task to the right model. - [Typed decision models](https://bradshaw.cloud/strategy/system-one-models/) — the cheap judgment layer in front of a frontier call. - [All strategy insights](https://bradshaw.cloud/strategy/)