Entities

An AI research and deployment company, and the maker of ChatGPT and the GPT model family. Its API is what a large share of other companies' AI products are built on.

The code hosting platform owned by Microsoft, and the home of Copilot — one of the earliest and most widely deployed coding assistants.

The Alphabet subsidiary whose DeepMind and Google Research divisions build the Gemini models. Its models reach readers through Search, Workspace and Android as well as through the Gemini apps directly.

An AI safety and research company, and the maker of the Claude model family and Claude Code. It serves its models as a hosted product and publishes research on model behaviour alongside them.

A Chinese AI lab known for releasing open-weight reasoning and coding models. Its releases are watched for what they show about training costs as much as for what the models can do.

The AI code editor made by Anysphere, and one of the most widely used AI-native development environments.

The parent company of Facebook, Instagram and WhatsApp, and of Meta AI's Muse assistant. Among the largest labs it is the most committed to open weights, publishing the Llama family for anyone to run.

The accelerated-computing company whose GPUs and DGX systems train and serve most frontier models. Its data-centre roadmap is a constraint the rest of the field plans around, because training capacity depends on it.

The device maker, whose Apple Intelligence features run models on-device and route the rest through a private cloud it operates itself.

The software company behind Windows, Azure and Copilot, and the largest investor in OpenAI. Azure resells models from several labs beside Microsoft's own, which makes it a route to market for its competitors as well as its partner.

The AI company behind Grok, founded by Elon Musk and operated alongside X. It distributes its assistant through X itself, which is unusual among the major labs.

A Chinese AI company publishing multimodal generation models spanning text, speech and video.

The workplace messaging platform owned by Salesforce.

The Chinese cloud and commerce group whose Qwen team publishes the Qwen model family. Qwen models are both served through Alibaba Cloud and released openly, which has made the family a common base for work done elsewhere.

A routing service that exposes many models from many providers behind one API, which makes it a common way to compare them.

The platform where open models, datasets and demos are published and downloaded. It is where most open-weight releases land, which makes it the reference point for what is available.

The enterprise CRM company, whose Agentforce line puts AI agents inside the products its customers already run.

A generative video company building camera-control models for AI footage.

A voice AI company building speech synthesis and dubbing models, used as much for audiobooks and localisation as for voice agents.

An AI search company building an answer engine and, more recently, a browser. It is one of the few search products built on models it does not train itself.

The cloud and commerce group, whose AWS division sells model hosting and its own training chips. AWS is where a large share of models are served, and its silicon is the main alternative to NVIDIA's.

No definition written; coverage recorded from the feed.

Anthropic's agentic coding tool, which runs in the terminal and the IDE. It is the agent most directly comparable to OpenAI's Codex, and the two are usually assessed against each other.

OpenAI's software-engineering agent, which runs in the terminal, the IDE and the cloud.

Google's multimodal model family and the assistant built on it. It is embedded across Google's products, which gives it a distribution no standalone assistant matches.

Anthropic's assistant and the model family behind it. It is used both as a consumer product and, through the API, as the base for a large share of business tooling.

xAI's assistant, distributed through X and through its own apps, which gives it a direct audience the other assistants have to build separately.

No definition written; coverage recorded from the feed.

Alibaba's model family, widely used as a base for open-weight fine-tunes.

OpenAI's assistant, and the product that made conversational AI mainstream. It remains the reference point most other assistants are compared against.

No definition written; coverage recorded from the feed.

Meta's AI assistant and the agent platform built around it, distributed through Meta's apps and, increasingly, as a desktop agent.

Zhipu AI's model family, published openly and served through its Z.ai platform.

No definition written; coverage recorded from the feed.

No definition written; coverage recorded from the feed.

No definition written; coverage recorded from the feed.

Moonshot AI's assistant and long-context model family.

No definition written; coverage recorded from the feed.

Microsoft's assistant, embedded across Windows, Office and GitHub — which makes it the assistant most office workers meet first.

NVIDIA's desktop AI computer, aimed at developers who want to run models locally rather than rent capacity.

No definition written; coverage recorded from the feed.

No definition written; coverage recorded from the feed.

No definition written; coverage recorded from the feed.

No definition written; coverage recorded from the feed.

ByteDance's video generation model family, distributed through its own products as well as by API.

No definition written; coverage recorded from the feed.

No definition written; coverage recorded from the feed.

An autonomous software-engineering agent from Cognition, and one of the earlier products to market an agent that runs a whole task rather than autocompleting lines.

The Model Context Protocol, an open standard for connecting models to tools and data sources. It has been adopted well beyond Anthropic and is implemented by a range of tools and providers.