← AI Terminology
Frontier Model
A frontier model is a leading general-purpose model at the cutting edge of capability (and often cost), typically from major labs, against which others are compared.
The set moves as new flagships release.
The set moves as new flagships release.
Why It Matters in AI
Product and policy debates orbit frontier systems: what they can automate, dual-use risks, and eval gaps. Routing easy traffic to cheaper models and hard traffic to frontier models is standard architecture.
Key Points
| Aspect | Description |
|---|---|
| Cost | Premium price; use via routers |
| Access | Usually API; weights often closed |
| Safety | Focus of evals and governance |
| Traits | Top benchmarks, multimodal, large context, heavy post-training |
| Related | SLM, test-time compute, scaling laws |
| Examples | Flagship GPT/Claude/Gemini/o-series class models |
Simple Analogy
The current world-record marathon runners of AI — not every jogger, only the athletes setting the pace of the field.
Common Usage Examples
- Router: frontier for hard reasoning only
- Compare mid-tier vs frontier on golden set
- Track price/perf as frontier moves
- Policy: extra review for frontier tool use
Summary
In short: Frontier models are the current capability leaders — the top-tier systems products and policymakers measure everyone else against.