The Frontier Model Landscape: Who Builds What
Closed, open-weight, reasoning, multimodal — a field guide to the major AI model families and how they differ.
Key takeaways
- A small number of labs — OpenAI, Anthropic, Google DeepMind, Meta, xAI and several Chinese and European labs — train frontier-scale models.
- Closed models lead on many benchmarks; open-weight models trail closely and dominate self-hosted and fine-tuning use cases.
- Mixture-of-Experts architectures let models grow total parameters without proportional compute per token.
- Reasoning (thinking) modes, long context, multimodality and tool use are now standard frontier features.
- Choose models by evaluating on your own tasks: quality, latency, cost per completed task, and deployment constraints.
Three years ago, "the AI model" usually meant one product. Today there is a crowded and fast-moving field of model families, each with different strengths, licenses and deployment options. This field guide maps the landscape and explains the technical dimensions that actually differentiate models.
The major model families
| Lab | Family | Access | Known for |
|---|---|---|---|
| OpenAI | GPT series, o-series reasoning models | Closed; API, ChatGPT, Azure | Consumer reach, early reasoning models, broad multimodality |
| Anthropic | Claude (Haiku, Sonnet, Opus, Fable) | Closed; API, Bedrock, Vertex AI, Foundry | Coding and agentic work, long context, safety research |
| Google DeepMind | Gemini; open Gemma models | Closed (Gemini) and open-weight (Gemma) | Native multimodality, very long context, TPU-scale training |
| Meta | Llama | Open weights (custom license) | Seeding the open ecosystem and fine-tuning community |
| xAI | Grok | Closed; some older weights released | Very large GPU clusters, integration with X |
| Mistral AI | Mistral, Mixtral, Codestral | Open-weight and commercial | Efficient European models, early open MoE |
| DeepSeek | DeepSeek-V3, R1 | Open weights (MIT for R1) | Training efficiency, open reasoning model |
| Alibaba | Qwen | Largely open weights | Broad size range, strong multilingual and coding variants |
Others matter too: Microsoft's Phi small models, Cohere's enterprise-focused Command models, Amazon's Nova family, and Chinese labs including Moonshot (Kimi) and Zhipu (GLM).
Closed vs open-weight
"Open" in AI usually means open weights: you can download the trained parameters and run them yourself, but the training data and full recipe are often not released. Licenses vary from permissive (Apache 2.0, MIT) to custom terms with usage restrictions.
- Closed models offer frontier capability, managed infrastructure, safety systems and enterprise support — but you depend on the vendor's API and pricing.
- Open-weight models offer control: on-premises deployment, data residency, deep fine-tuning and no per-token vendor fee. You take on serving costs and safety responsibilities.
The capability gap between the best closed and best open models has fluctuated but generally stayed in the range of months, not years — a major reason open models are widely used for cost-sensitive and private workloads.
Dense vs Mixture-of-Experts
In a dense model, every parameter participates in processing every token. In a Mixture-of-Experts (MoE) model, each Transformer layer contains many "expert" feed-forward networks, and a learned router sends each token to only a few of them.
DeepSeek-V3, for example, has 671 billion total parameters but activates about 37 billion per token. MoE lets labs scale knowledge capacity cheaply in FLOPs — but every expert still has to sit in memory, which makes serving MoE models a memory and networking challenge.
The features that now define "frontier"
- Reasoning / thinking — models generate intermediate reasoning before answering, with controllable effort.
- Long context — hundreds of thousands to a million or more tokens, enough for entire codebases or document sets.
- Multimodality — image and document understanding are standard; audio and video input are common; some models generate images and speech natively.
- Tool use — reliable function calling, web search, code execution and computer use, which make agents possible.
- Structured outputs — guaranteed JSON conforming to a schema.
How to read benchmarks
Benchmarks are useful but easy to over-interpret. Common ones include SWE-bench Verified (fixing real GitHub issues), GPQA Diamond (graduate-level science questions), Humanity's Last Exam (very hard expert questions), various agentic benchmarks such as Terminal-Bench and τ-bench, and the crowd-sourced LMArena leaderboard.
Choosing a model: a practical checklist
- Build a small eval set from your real tasks — 50 to 200 examples with clear grading.
- Measure cost per completed task, not per token. A cheaper model that needs more retries may cost more.
- Check latency — time to first token and total time, at realistic output lengths.
- Consider deployment — data residency, cloud marketplace availability, compliance, private networking.
- Try effort and size tiers — the strongest model at low effort sometimes beats a smaller model at high effort.
- Re-evaluate regularly — the leaderboard shifts every few months.
Where the landscape is heading
The frontier is shifting from single-turn quality toward long-horizon reliability: how well a model completes a task that takes hundreds of tool calls over hours. At the same time, small models keep getting better, making on-device and edge AI practical. Expect continued pressure on price per token, more specialization (coding, science, voice), and more competition from open-weight releases.