Who Makes AI Models? OpenAI, Anthropic, Google, DeepSeek

By the WahLiao desk · Last verified 3 October 2026

Super Intelligence, the kind of AI behind ChatGPT, Claude and Gemini, comes from a fairly small group of companies. In 2026 three American firms are setting the pace: OpenAI (ChatGPT, which runs on its GPT-6 family), Anthropic (Claude, whose Opus 5.5 and Sonnet 5.5 arrived in September 2026) and Google DeepMind (Gemini, whose newest general release is Gemini 3.8 Flash, with Gemini 4 Argon announced on 30 September 2026). Around them you’ll find Meta, Microsoft and xAI in the US, Mistral in France, DeepSeek and Alibaba in China, and Singapore’s very own SEA-LION from AI Singapore. Almost all of them work from Singapore today, and several let anyone download their models for free.

Here’s the one rule we’d keep in your pocket: the brand on the app isn’t always the maker of the model inside it. Microsoft’s Copilot runs on OpenAI models as well as Microsoft’s own, and Singapore’s SEA-LION has been built on other makers’ models, from Meta’s Llama and Alibaba’s Qwen to, most recently, NVIDIA’s Nemotron. So before you ask which one is “best”, ask who made the model and where your data goes.

Who makes AI models: quick facts

US makers	OpenAI, Anthropic, Google DeepMind, Meta, Microsoft, xAI (now part of SpaceX)
European maker	Mistral AI, Paris
Chinese makers	DeepSeek (Hangzhou), Alibaba (Qwen)
Singapore’s own	SEA-LION, by AI Singapore, tuned for Southeast Asian languages
Open-weight leaders	DeepSeek, Mistral, Qwen, Google Gemma, OpenAI gpt-oss, SEA-LION
Chip layer	Nvidia dominates; Google TPUs and Amazon Trainium are the main alternatives
Release pace	New versions every few weeks; the names in this guide are as of 3 October 2026

The WahLiao Verdict

Big three	ChatGPT, Claude and Gemini all officially support Singapore.
Best value	Start free. Pay only when a limit starts costing you time.
Watch the data	DeepSeek says it stores personal data in China. Keep that in mind when you choose.
Open-weight	Great for developers and privacy-minded firms; most of us won’t need it.
Version numbers	Don’t chase every release. The family matters more than the point number.

The main makers at a glance

Model names change quickly, so this table gives each family plus a recent release we could confirm from the maker or from reputable reporting. “Open-weight” means you can download the trained model and run it on your own computers; “closed” means you can only reach it through the maker’s app or a paid connection.

Maker (home)	Chatbot and model family	Open or closed
OpenAI (US)	ChatGPT; GPT-6 family: Astra (early September 2026), Sol and Luna (22 September) and GPT-6.1 Sol (29 September 2026)	Closed flagship; open-weight gpt-oss models since August 2025
Anthropic (US)	Claude; Opus 5.5 (22 Sept 2026), Sonnet 5.5 (28 Sept 2026), plus the higher-end Fable and Mythos 5.1 (1 Sept 2026)	Closed
Google DeepMind (UK/US, part of Alphabet)	Gemini; Gemini 3.8 Flash (2 September 2026); Gemini 4 Argon announced 30 September 2026, rolling out first to a small group of trusted cyber defenders, with paid developers and Google AI Ultra subscribers to follow	Closed Gemini; open-weight Gemma 4 (from March 2026)
Meta (US)	Meta AI in WhatsApp, Instagram and Facebook; Muse Spark (April 2026), updated to 1.3 in September	Muse Spark closed; older Llama models open-weight
Microsoft (US)	Copilot; OpenAI models, its own MAI models and, in Microsoft 365 Copilot, Anthropic’s Claude as an option	Closed
xAI (US, part of SpaceX since February 2026)	Grok, inside X and on its own app; Grok 4 series, most recently Grok 4.7 (September 2026)	Closed
Mistral AI (France)	Le Chat; Mistral Medium 3.5 and Mistral Large 3	Mostly open-weight
DeepSeek (China)	DeepSeek app; DeepSeek-V4 Pro and Flash (full release July to August 2026)	Open-weight, MIT licence
Alibaba (China)	Qwen Chat; Qwen3.8-Max (August 2026)	Open-weight: Max under its own licence with limits on large commercial use; smaller Qwen3.8 models under Apache 2.0
AI Singapore (Singapore)	SEA-LION; Nemotron-SEA-LION-v4.8 (18 September 2026), built with NVIDIA, after Qwen-SEA-LION-v4.5 (May 2026)	Open-weight

A word on availability. OpenAI, Anthropic and Google all list Singapore as a supported country for their chatbots, and none of the three lists mainland China. Anthropic says it’ll open a Singapore office in October 2026, its first in Southeast Asia; OpenAI opened one here in late 2024. You can generally reach the others on the web or through their apps, but do check each service’s terms before you lean on it for work.

Closed vs open-weight: why it matters

A closed model lives on the maker’s servers. You type, it answers, and the company keeps control of the model, the safety rules and your data under its privacy terms. ChatGPT, Claude, Gemini, Grok and Meta’s Muse Spark all work this way. For most of us it’s simply the easiest option: nothing to install, and the newest version just shows up.

An open-weight model is one whose trained “weights”, the billions of numbers that make up the model, are published for anyone to download. DeepSeek-V4 and Qwen-SEA-LION-v4.5 use the MIT licence; Google’s Gemma 4 and Mistral Large 3 use Apache 2.0. A company can run a model like this on its own machines, so customer data never leaves the building, and it can adjust the model for a particular language or job. That’s exactly what AI Singapore did. Its Qwen-SEA-LION-v4.5, released in May 2026, is built on Alibaba’s open-weight Qwen3.6 and tuned for Burmese, Filipino, Indonesian, Malay, Tamil, Thai and Vietnamese as well as English. Its newest release, Nemotron-SEA-LION-v4.8, announced on 18 September 2026, was built with NVIDIA on its Nemotron models. Earlier SEA-LION versions were built on Meta’s Llama.

Two gentle cautions. Open-weight isn’t quite the same as “open source” in the full sense, because the training data and methods usually stay private. And running a Chinese model’s downloaded weights on your own computer is a different matter from using its app: DeepSeek’s privacy policy says it “collects, processes and stores your personal data in the People’s Republic of China”. Our guide to what happens to what you type walks through the settings and the PDPA angle.

The line keeps moving, too. Meta made its name with open Llama models but kept its newest model, Muse Spark, closed, while saying it may release an open version later. Alibaba went the other way: in August 2026 it published the weights of its top Qwen3.8-Max, under a licence that limits large commercial use. And OpenAI, closed for a long time, released open-weight gpt-oss models in August 2025.

The chip layer: who supplies the engines

Every one of these models is trained and run on specialised chips in large data centres, and most of those chips come from one company. Nvidia reported revenue of US$96.2 billion for the quarter ended 26 July 2026, of which US$89.0 billion came from its data centre business, up 117% on a year earlier. As its chief executive, Jensen Huang, put it: “compute is revenue”.

The main alternatives come from the cloud giants themselves. Google designs its own TPU chips, and Anthropic agreed in October 2025 to use up to one million of them, while also training on Amazon’s Trainium chips and Nvidia GPUs. Spreading the work across several chip makers is now a common way for labs to keep costs in check and avoid shortages. For how this plays out in electricity and land here at home, see energy, water and data centres in Singapore. Chip company results swing about a lot, and nothing on this page is a view on any share.

Which one should I use?

We don’t have a favourite, and for everyday tasks the leading chatbots are closer than their marketing suggests. Here’s a practical way to choose.

Try the free tiers of two or three with a real task from your week, like a letter, a spreadsheet question or a school worksheet, and keep the one whose answers you find yourself checking least. Match the tool to where you already work: Gemini sits inside Google’s apps, Copilot inside Microsoft 365, and Meta AI inside WhatsApp and Instagram where it’s offered. Pay only when a limit costs you time, and compare plans in S$ on our free vs paid plans page. Singapore citizens aged 18 and above who finish an eligible SkillsFuture AI course can get six months of a premium tool free. Whichever you pick, they all make confident mistakes, so do check anything that matters, and have a read of using it well for the habits that help most. For Singapore’s wider role, from AI Singapore to AI Verify, see Singapore’s own AI story.

Model news moves fast, and this page may slip behind between our checks. If you spot something we’ve got wrong, tell us and we’ll fix it.

Who makes AI models: FAQ

Who owns ChatGPT?

ChatGPT is made by OpenAI, which is based in San Francisco. Microsoft is a major partner and investor and uses OpenAI models in Copilot, but it doesn’t run ChatGPT.

Is Gemini the same as Google DeepMind?

Not quite. Gemini is the model family and the chatbot; Google DeepMind is the Alphabet research unit that builds it. Google also publishes a separate open-weight family called Gemma.

Is DeepSeek safe to use in Singapore?

If you use its app, bear in mind that DeepSeek’s privacy policy says personal data is stored in China. Keep sensitive, personal and work information out of it, as you would with any chatbot. Its open-weight models can also run on your own hardware, which keeps your data local.

What is SEA-LION?

SEA-LION is a family of open-weight language models from AI Singapore, the national programme, tuned for Southeast Asian languages. The newest release, Nemotron-SEA-LION-v4.8, announced on 18 September 2026, was built with NVIDIA; the earlier Qwen-SEA-LION-v4.5, released in May 2026, is built on Alibaba’s Qwen3.6.

What does open-weight AI mean?

It means the trained model file is published, so anyone can download, run and adapt it within the licence. It usually doesn’t include the training data, so it’s more open than a closed model but not fully “open source”.