Frontier model

"Frontier model" is the term of art for whatever sits at the capability ceiling at a given moment. It's a moving target. In 2023, GPT-4 was frontier. In 2024, Claude 3.5 Sonnet and GPT-4o joined it. In 2026, the frontier is GPT-5, Claude Opus 4.7, Gemini 3 Pro, and depending on who you ask, DeepSeek V4 or Grok 4 — models capable of frontier-level reasoning, coding, multimodal understanding, and agentic tool use. Frontier is defined by capability, not size. A 2T-parameter dense model trained on noisy data is not frontier. A 400B-parameter MoE with state-of-the-art post-training, RLHF, and reasoning fine-tuning is. Open-weights releases (DeepSeek V4, Llama 4 405B) have repeatedly approached or matched closed-frontier capabilities, which has scrambled what "frontier" means commercially even when capabilities are similar. The label matters because it tracks what's possible right now versus what was possible a year ago. A task that frontier models in 2024 couldn't do reliably (multi-step agentic coding, long-document reasoning, complex math) is now routine on Claude Opus 4.7 or GPT-5 Thinking. If you tried something with GPT-4 in 2024 and gave up, it's probably worth retrying with a 2026 frontier model — the gap is enormous. Conversely, if a non-frontier model is cheaper and "good enough," frontier is often overkill.