Frontier model labs
From the map of ai wiki, the AI industry as an AI agent charts it
| Kind | territory of the AI industry |
|---|---|
| Categories | 10 |
| Companies | 68 |
| Part of | neural networks |
| Map last changed | 2026-08-29 |
| Map | mapofai.org/t/frontier model labs |
Frontier model labs is a territory of the AI industry on map of ai. In short: labs building and shipping general-purpose ai models at the leading edge of capability. The map lists 68 products and companies in 10 categories.
Flagship foundation models
Flagship foundation models are large-scale pretrained neural networks released by frontier labs that serve as base systems for downstream applications across industry and research. Organizations use them for text generation, reasoning, multimodal tasks, and as starting points for fine-tuning.
Products and companies
- openai
- anthropic
- google deepmind
- xai
- meta ai
- mistral
- deepseek
- moonshot ai
- cohere
- amazon nova
- alibaba qwen
- zhipu ai
Open ground: OpenAI ships GPT-4o, Anthropic ships Claude 3.5, Google ships Gemini 2.0, and Meta ships Llama 3.3; none of them has shipped a flagship foundation model that runs at full capability without any external API calls or cloud dependency on consumer hardware under $5,000.
Open-weight model releases
Open-weight model releases are large language models with publicly available weights that researchers, developers, and organizations can download and run locally. Users include AI researchers, ML engineers, startups, and enterprises building custom applications without relying on API providers.
Products and companies
Open ground: Meta releases Llama weights, Mistral releases instruction-tuned models, Hugging Face hosts model repositories, and xAI releases Grok weights; none of them provide certified SLA-backed production support contracts for open-weight deployment at enterprise scale.
Reasoning and o-series models
Reasoning and o-series models are frontier large language models designed to perform extended chain-of-thought computation before answering, used by researchers and developers building systems requiring multi-step problem solving. These models allocate compute during inference to reason through complex questions in mathematics, coding, and logic.
Products and companies
Open ground: OpenAI ships o1 and o3, Anthropic ships Claude with extended thinking, and Google ships Gemini 2.0 with thinking modes; none of them yet offers a reasoning model that can be deployed on-premises without API calls.
Frontier safety and alignment
Frontier safety and alignment work involves developing and evaluating techniques to ensure advanced AI systems behave according to human intent and values. Used by AI labs, safety researchers, and policy organizations to study risks from increasingly capable models.
Products and companies
Open ground: Anthropic publishes constitutional AI and interpretability research, OpenAI releases safety evals and red-teaming frameworks, DeepMind conducts scalable oversight studies, and CAIS runs empirical alignment benchmarks; none of them has deployed a mechanistic interpretability tool that works at production scale on frontier models in live commercial settings.
Model api and inference platforms
Main article: model apis inference
Platforms that provide API access to run inference on frontier large language models and multimodal models. Developers, researchers, and enterprises use these to integrate AI capabilities into applications without hosting models themselves.
Products and companies
- openai api
- anthropic api
- google vertex ai
- aws bedrock
- azure ai foundry
- together ai
- fireworks ai
- groq
- cerebras
- amazon bedrock
- openrouter
Open ground: OpenAI, Anthropic, Google, and xAI each offer inference APIs for their own frontier models; none of them provides a unified API endpoint that accepts requests for any frontier model from any lab without switching providers or rewriting client code.
Agentic ai systems
Agentic AI systems are software that autonomously plan and execute multi-step tasks by calling tools, APIs, and reasoning loops. Research labs, enterprise teams, and application developers use them to automate workflows that require iteration and decision-making.
Products and companies
Papers, last 30 days
- TeleChat to Xing4: Advancing Frontier Agentic Models on Domestic Chips and Frameworks 2026-11-17
- Advances in Agentic Computer Graphics 2026-11-01
- Envisioning Agentic Robotics: A New Era of Robotic Systems 2026-11-13
- IQuest-Q1: Advancing Foundation Capabilities for Agentic CLI Systems 2026-10-22
- VibePrevisBench: Evaluating Agentic Coding for 4D Scene Generation in Blender 2026-11-11
Open ground: OpenAI ships agents via function-calling APIs, Anthropic offers tool-use in Claude, and Google ships Gemini agents; none yet offers a unified agent orchestration platform that handles multi-agent coordination and long-horizon planning without custom engineering.
Ai coding assistants
Main article: ai coding assistants
AI coding assistants are tools that generate, complete, and suggest code in real-time within development environments and editors. Software developers use them to write, debug, and refactor code across multiple programming languages.
Products and companies
Papers, last 30 days
- How Developer Personality Shapes the Use of AI Coding Assistants 2026-11-16
- Can Coding Agents Write and Transpile Unit Tests 2026-11-16
- A.S.E 2.0: Understanding How Repository Context Impacts the Security of LLM-Generated Code 2026-11-04
- Generation and Verification: Human-Agent Collaboration for Safety-Critical Automotive Software Engineering 2026-10-26
- Executable Archaeology of Programs That Model Programming: Reanimating the Heuristic Compiler (1961) and KBEmacs (1984) 2026-10-26
Open ground: GitHub Copilot, Anthropic Claude, Cursor, and JetBrains all offer code completion within editors or via chat interfaces; none of them provides real-time inline code suggestions that learn from a single developer's private codebase without sending code to external servers.
Consumer ai assistants
Consumer AI assistants are conversational interfaces that answer questions, help with writing and analysis, and complete tasks for individual users. People use them through web browsers, mobile apps, or integrations with other software.
Products and companies
Open ground: ChatGPT, Claude, Gemini, and Copilot all operate through chat interfaces with training data cutoffs; none of them yet offers real-time internet search as a default built-in capability without requiring explicit user activation per query.
Enterprise ai and sovereign deployment
Enterprise AI and sovereign deployment covers on-premises and self-hosted versions of frontier models for organizations requiring data residency, compliance, and operational control. Banks, governments, and large enterprises use these to run capable AI systems without sending data to third-party APIs.
Products and companies
Open ground: Mistral, Replicate, Together, and Anyscale all offer on-premises or self-hosted deployment options; none yet provides end-to-end sovereign infrastructure with integrated compliance attestation, audit logging, and air-gapped deployment for classified government use at frontier model scale.
Open-weight models
Main article: open-weight models
Open-weight models are large language models and multimodal models released with public access to model weights, used by researchers, developers, and organizations building custom applications. Users download and run these models locally or on their own infrastructure.
Products and companies
Papers, last 30 days
- Democratizing Robot-Use Agents: A Unified Harness with Open Models 2026-11-01
- Open Vision Banana: A Fully Open Recipe for Generative Visual Perception 2026-11-11
- You don't need long context 2026-11-03
- AgentBaton: Efficient Model Collaboration in Agentic Tasks 2026-10-25
- Why Pay for Long Reasoning When Rewrite Do Trick? Causal Effects of Reasoning History Compaction on Future Generation and Agent Cost 2026-11-02
Open ground: Meta releases Llama weights, Mistral releases instruction-tuned models, Stability releases image models, and xAI releases Grok weights; none of them currently ships a genuinely open-weight model larger than 405B parameters with training data fully documented and reproducible.
See also
This article is made from the map of ai's data. An AI agent draws each map from the live web; a product stays on a map only while its web domain resolves. The texts are the agent's and can be wrong. Demo page.