AI companies: who actually builds the AI models you use in 2026?

The AI tools you use, from ChatGPT to Gemini, are built by a few dozen companies, mostly in the United States and China, with a handful of players in Europe, Canada, Israel, Japan and South Korea. This guide explains who does what, who leads the ranking as of October 22, 2026, and why those differences matter to you.

Maker, model, assistant: three different things

Before touring the companies, a bit of vocabulary will save you a lot of confusion. An AI model is the “engine”: a program trained on huge amounts of text and images that can generate an answer. The maker is the company that designs and trains that model. The AI assistant is the app you actually chat with.

The three often overlap. Anthropic makes the Claude models and the Claude assistant; OpenAI makes the GPT models and ChatGPT. But not always. Perplexity, Duck.ai and Poe offer assistants that give you access to models from several makers. And Microsoft puts OpenAI models inside Copilot while also building its own.

A maker can ship its models in three ways:

• in its own assistant, for the general public;
• through an API, a technical gateway that developers plug into their software, billed by usage;
• as a download, when it publishes the model’s “weights”, the internal settings that let anyone run the model on their own server.

That last point splits makers into three camps: closed, open and mixed. More on that below.

The specialist US labs

Three American companies were founded specifically to build AI models, and that is almost all they do.

OpenAI, founded in 2015 in San Francisco as a nonprofit, brought conversational AI to the mainstream when it launched ChatGPT on November 30, 2022. According to its profile on this site, it restructured in October 2025: a foundation now controls a public benefit corporation, OpenAI Group PBC. Its flagship GPT models are closed, but it also releases an open-weights family called gpt-oss.

Anthropic, founded in 2021 by Dario and Daniela Amodei, makes the Claude models, offered in Haiku, Sonnet, Opus and Fable tiers, from fastest to most capable. It is also a public benefit corporation, known for its research on model safety. No Claude model can be downloaded.

xAI, launched by Elon Musk in July 2023, builds the Grok models and assistant. According to its profile, SpaceX announced its acquisition of xAI on February 2, 2026, and the company kept its brand names. Its recent models are closed.

The tech giants that build their own models

Other makers are huge groups for which AI is one business among many. Their edge is distribution: their models go straight into products people use every day.

• Google DeepMind was formed in 2023 by merging DeepMind, founded in London in 2010, with Google Brain, whose researchers published “Attention Is All You Need” in 2017, the paper that introduced the Transformer. It makes the closed Gemini models and the open Gemma models.
• Meta, the parent company of Facebook, Instagram and WhatsApp, long bet on its open Llama models. Since April 2026, its new models are called Muse: the flagship, Muse Spark, is closed, but Muse Spark Open can be downloaded.
• Microsoft builds the small open Phi models and the MAI family, while remaining OpenAI’s main partner.
• Amazon makes the closed Nova models, sold to businesses through its AWS cloud.
• NVIDIA, the leading supplier of the GPUs that power AI, releases its Nemotron models with their weights and even their training data.
• IBM publishes the Granite family, open and designed for business use.

The Chinese labs

China is the second major hub. What sets it apart: most of its makers release their large models with open weights, often under very permissive licenses such as MIT.

• DeepSeek, based in Hangzhou, became known for reasoning models released freely. Its V4 models are published under the MIT license.
• Alibaba Cloud develops the Qwen family, one of the most widely used by developers, with open models in every size and more powerful models kept for its API.
• Moonshot AI, founded in Beijing in 2023, makes the Kimi assistant and the Kimi K models, open under a modified MIT license.
• Zhipu AI, spun out of Tsinghua University and known internationally as Z.ai, publishes the GLM models under the MIT license.
• Others follow: MiniMax, ByteDance (TikTok’s parent company) with its Seed team, Tencent with Hunyuan, Xiaomi with MiMo, Baidu with ERNIE, as well as StepFun and inclusionAI (Ant Group).

Europe and the rest of the world

Outside the two main hubs, makers are fewer and often more specialized.

Mistral AI, founded in Paris in April 2023 by Arthur Mensch, Guillaume Lample and Timothée Lacroix, is Europe’s leading model maker. It runs the Le Chat assistant, renamed Mistral Vibe in 2026, and releases some of its models under the Apache 2.0 license. Its pitch: a company bound by the GDPR, with some models you can run on your own servers. Europe’s place in this race is covered in our article on European AI versus the US and China.

Elsewhere you will find Cohere in Canada, focused on business models and document search; AI21 Labs in Israel, with its Jamba models; Sakana AI in Japan, co-founded by one of the authors of the original Transformer paper; and Upstage in South Korea, with its Solar models. The full list, with a profile for each company, is on our AI labs page.

Maker	Country	Weights policy	Models	Consumer assistant
OpenAI	United States	Mixed	GPT, gpt-oss	ChatGPT
Anthropic	United States	Closed	Claude	Claude
Google DeepMind	United Kingdom (Google)	Mixed	Gemini, Gemma	Gemini
Meta	United States	Mixed	Muse, Llama	Meta AI
xAI	United States	Closed	Grok	Grok
Mistral AI	France	Mixed	Mistral, Ministral, Codestral	Le Chat (Mistral Vibe)
DeepSeek	China	Open	DeepSeek V and R	DeepSeek
Alibaba Cloud	China	Mixed	Qwen	Qwen Chat
Moonshot AI	China	Open	Kimi K	Kimi
Zhipu AI	China	Open	GLM	Z.ai

Weights policy and country: maker profiles on this site, data as of October 22, 2026. Google DeepMind is listed under the United Kingdom, where its original headquarters is.

Who leads the ranking?

To compare makers, people usually look at their best models. Our ranking is based on results published by Epoch AI, an independent research institute, across several benchmarks: sets of tests in math, science, programming and agent tasks. The score out of 100 is the average of those tests. Here is the top of the ranking as of October 22, 2026:

Rank	Model	Organization	Index
1	Claude Sonnet 5.5	Anthropic	95.9
2	Claude Opus 5.5	Anthropic	95.9
3	Gemini 4 Argon	Google DeepMind	95.3
4	GPT-6 Astra	OpenAI	94.7
5	Claude Opus 5	Anthropic	92.7
6	Claude Fable 5.1	Anthropic	92.6
7	GPT-6.1 Sol	OpenAI	92.1
8	Claude Fable 5	Anthropic	90.8
9	GPT-6 Sol	OpenAI	89.2
10	GPT-5.5	OpenAI	89

The ranking tells you something important: at the very top, the race is between a handful of US labs, separated by a few points. It doesn’t tell you everything, though. A model can ace the tests and still be too expensive for your needs, or perform less well in your language. To learn how to read these numbers without being misled, see our article on AI benchmarks and their pitfalls. The full model ranking is updated every week.

One flagship model per maker, side by side

The ranking only shows the very top. This table puts a recent model from eight makers side by side, with their developer pricing, their context window (how much text they can read at once) and whether their weights are open.

Model	Developer	Release	Context (tokens)	Input ($/M tokens)	Output ($/M tokens)	Open weights
Claude Sonnet 5.5	Anthropic	September 28, 2026	1 M	$2	$10	No
GPT-6 Astra	OpenAI	September 4, 2026	1.05 M	$10	$50	No
Gemini 4 Argon	Google DeepMind	October 13, 2026	1.05 M	$2	$10	No
Grok 4.7	xAI	September 21, 2026	500 k	$2	$6	No
Mistral Large 4	Mistral AI	October 6, 2026	1.05 M	$1.36	$4.18	No
DeepSeek V4.1 Flash 1017	DeepSeek	October 17, 2026	1.05 M	$0.15	$0.60	Yes
Kimi K3	Moonshot AI	July 16, 2026	1.05 M	$3	$15	Yes
Muse Spark Open	Meta	October 15, 2026	1.05 M	$1.25	$4.25	Yes

Prices are in US dollars per million tokens, the small chunks of words a model processes. They apply to the API, not to assistant subscriptions. Two worlds quickly appear: closed US models among the strongest performers, and open models, often Chinese, that are much cheaper to run. To compare other models on your own criteria, use the model comparison tool.

Closed, open or mixed: why it matters

The weights policy is probably the most practical difference between makers.

• Closed (Anthropic, xAI, Amazon): you can only reach the model through the maker’s assistant or API. Your data goes through its servers, under its rules.
• Open (DeepSeek, Moonshot AI, Zhipu AI, NVIDIA, IBM): the open weights can be downloaded. A company can run the model in-house or with a host of its choice, and adapt it. Large models do, however, require powerful hardware.
• Mixed (OpenAI, Google, Meta, Mistral AI, Alibaba): the most powerful models stay closed, while lighter models are released.

Be careful with the term “open source”: most of the time only the weights are published, not the training data or code, and each license has its own conditions. Our article on open-source versus proprietary AI explains the nuances.

Check yourself You want to use a Chinese model without your documents passing through servers in China. Is that possible?

Yes, as long as you choose an open-weights model, such as those from DeepSeek or Zhipu AI, and run it on your own servers or with a host located where you want. If you use the maker’s own assistant or API, however, your data is processed by that maker under its privacy policy.

What the big-picture numbers say

Beyond individual models, a few public datasets show how much weight each region carries.

Year	United States	China	United Kingdom	South Korea	Other
2023	81	23	2	1	12
2024	55	30	1	1	11
2025	62	34	1	8	2

Number of notable AI models by country of origin, according to Epoch AI (retrieved October 22, 2026). “Other” includes Canada.

These figures confirm US dominance in both money and model count, and China’s strong position despite much lower private investment. They say nothing, however, about public funding or the actual cost of training each model.

All this growth also has a physical cost: data centers, electricity, water. We cover that in our article on how much energy AI uses.

How to choose between these companies

You don’t need to know every maker to choose well. Four questions are enough.

1. What will you do with it? For chatting, writing or searching, compare assistants rather than makers: one assistant can use several models.
2. Where does your data go? The maker’s country and privacy policy matter, especially for work documents. If you work with EU customers, the European regulation also applies: our summary of the EU AI Act and what it means for you covers the essentials.
3. Do you need to host the model yourself? If so, only open or mixed makers will do.
4. What is your budget? For developers, price gaps between models of similar performance can be huge.