Foundation Model Landscape

The frontier model landscape changes month to month, but the structure of it is more stable than the leaderboard suggests. Knowing the major families — and what each is optimized for — is more useful than chasing the latest benchmark.

The Major Frontier Labs

• Anthropic — Claude family. Strong reasoning, long context, character and safety work treated as first-class.
• OpenAI — GPT family plus reasoning models (o-series). Broadest tooling and ecosystem.
• Google DeepMind — Gemini family. Native multimodality, deep integration with Google products.
• Meta — Llama family. The dominant open-weights line; the foundation most fine-tuned models build on.
• Mistral — efficient open and commercial models, strong European footprint.
• xAI — Grok family.

Open vs Closed Weights

• Closed weights — accessed only via API. Easier to ship, no infrastructure burden, but you're a tenant of someone else's model.
• Open weights — Llama, Mistral, Qwen, DeepSeek. You can host, fine-tune, audit, and run offline. Lower per-token cost at scale; higher operational cost overall.

The right answer depends on your throughput, regulatory posture, and whether the marginal capability gap is worth the operational tax.

Capability Tiers

A rough mental model of the current tiers:

• Frontier — the largest, most capable models from each lab. Highest cost per token, best at hard reasoning.
• Mid-tier / fast — smaller, faster, cheaper variants of the frontier. The right default for most product features.
• Small / on-device — 1B–8B parameter models. Run on a laptop or phone, suitable for narrow tasks and latency-critical paths.

Most production systems route between tiers, sending only the requests that actually need frontier capability to the expensive model.

Specialized Models

• Reasoning models — trained to think through problems with explicit intermediate steps. Better on math, coding, and multi-step planning.
• Code models — tuned specifically for code generation and editing.
• Multimodal models — accept and produce mixes of text, images, audio, and video.
• Embedding models — produce vectors, not text. Different family entirely.

What to Watch

The interesting axes today aren't raw quality — they're context length, cost per token, latency, tool use reliability, and how well the model handles long-horizon agentic work. Those determine what you can actually build, not the leaderboard score.