A handful of organisations build the AI systems that most people use or read about. This guide sets out who they are, where their money comes from, and which model each one currently presents as its best. It reflects what the companies’ own pages said on 12 November 2026. Names change quickly in this field, so treat the list as a snapshot.

What an AI lab makes

An AI model is a piece of software trained on very large amounts of text, images and other data so that it can answer questions, write, analyse documents and produce computer code. The companies that train the largest ones are usually called labs. Each sells a family of models at different sizes and prices. The flagship is the most capable member of the family, and usually the most expensive to use.

Training these models needs vast data centres, which is why funding matters. A lab is either part of a large technology company that pays the bills from its other businesses, or it raises money from outside investors.

The current flagship models

LabFlagship modelDate given by the labWho can use it
AnthropicClaude Fable 5.29 November 2026Generally available
OpenAIGPT-6 AstraNot given on model pageListed for developers
Google DeepMindGemini 4 Carbon28 October 2026Selected cyber defenders only
MetaMuse Wonder12 October 2026Meta Model API and Muse Code
xAIGrok 4.7September 2026Listed for developers
MistralMistral Large 46 October 2026Public preview, weights since published
DeepSeekDeepSeek V4 ProNot given on pricing pageListed for developers
Flagship models as described on each lab’s own pages, 12 November 2026

Two things stand out. Five of the seven flagships carry a date from September 2026 or later, so any ranking goes out of date within weeks. And “flagship” does not always mean “available to you”.

The independent labs: OpenAI and Anthropic

OpenAI makes ChatGPT. Its model list names four flagship models and describes GPT-6 Astra as the one for “the most demanding work”. Microsoft has invested $13bn since 2019, according to Reuters, which named SoftBank, Amazon and Nvidia as investors in OpenAI’s most recent funding round. Reuters put that round at $110bn and the company’s valuation at $840bn. The Associated Press gave slightly higher figures of $122bn and $852bn. A valuation is the price that investors’ terms imply for the whole company. In June OpenAI said it had confidentially filed for an initial public offering (IPO), the first sale of a company’s shares on a stock market, without setting a timetable.

Anthropic makes Claude. Its models page lists Claude Fable 5.2 for the most demanding reasoning tasks and recommends the cheaper Claude Opus 5.5 for most work. Fable 5.2 was released on 9 November alongside Claude Mythos 5.2, which Anthropic says is the same underlying model without Fable’s safeguards on cybersecurity and biology work, offered only to vetted organisations. In May the company raised $65bn at a $965bn valuation, the Associated Press reported. On 10 November it began marketing an IPO of 825 million shares at an expected $70 to $78 each on the Nasdaq exchange. The sale had not been completed when that announcement was made.

The labs inside larger companies: Google DeepMind, Meta and xAI

Google DeepMind is Google’s AI division, so its funding comes from within Google. Its senior vice-president announced Gemini 4 Carbon on 28 October, calling it “our strongest model”. Access began with a limited group of cyber defenders, and Google gave no date for wider release.

Meta owns Facebook, Instagram and WhatsApp. It released Muse Wonder and Muse Wonder Full on 12 October and calls them its most capable models. Meta pays for this from its advertising business. Its third-quarter results forecast capital spending of $135bn to $145bn for 2026. That figure covers all of Meta’s capital spending, not only AI, although the company says it expects to keep investing heavily in AI.

xAI makes Grok. Its developer pages describe Grok 4.7 as “the most capable model we’ve built”, and its release notes list the model under September. xAI is no longer a stand-alone company. SpaceX, Elon Musk’s rocket business, acquired it in 2026 and then raised $75bn in a June IPO that valued the combined company at about $1.75tn, NPR reported.

A lab is either part of a large technology company that pays the bills from its other businesses, or it raises money from outside investors.”

The challengers: Mistral and DeepSeek

Mistral AI is based in Paris. It released Mistral Large 4 in preview on 6 October, describing it as its largest and most capable model, and published the weights later that month. Weights are the numbers a model learns in training. A lab that publishes them, an approach known as open weights, lets other people download and run the model on their own computers, subject to the licence. Mistral’s latest funding round raised €3bn at a valuation of more than €21bn. Samsung Electronics led it.

DeepSeek is a Chinese lab. Its pricing page lists two current models, V4 Pro and the cheaper V4.1 Flash, and it published the Flash model’s weights on 17 October. DeepSeek is a subsidiary of High-Flyer, a hedge fund. In June it raised more than $7.4bn at a valuation above $50bn, according to press reports summarised by SiliconANGLE. Those reports relied on unnamed sources, so the figures carry less weight than the others here.

Reported valuations of AI labs or their ownersUS dollars, billions
Reported valuations of AI labs or their owners
ItemUS dollars, billions
SpaceX, owner of xAI (June 2026)$1.75tn
Anthropic (May 2026)$965bn
OpenAI (as reported by Reuters)$840bn
DeepSeek (reported, June 2026)Above $50bn

Valuations come from different deals and dates. SpaceX owns xAI. DeepSeek’s figure is from press reports.Source: Associated Press, Reuters, NPR and SiliconANGLE. Chart: The Sidelight.

What these figures do and do not show

These numbers are not like for like. A private valuation is set by a small group of investors in one deal. A stock market value changes every day. Meta’s figure is a spending plan, not a company value. Some of the figures come from news agencies because the companies did not publish them.

This list is also not complete. Other companies train large models, and this guide covers only the seven named above.

How to read the next announcement

  • Check availability. Gemini 4 Carbon and Claude Mythos 5.2 show that a lab’s strongest model may be restricted to approved users.
  • Treat “most capable” as the lab’s own claim. Several labs here use similar words about their newest model.
  • Look at who is paying. A lab inside a larger company depends on that company’s budget. One funded by investors, or about to list its shares, depends on their continued support.
  • Expect the names to change. Most of the models in the table carry a release date from the past three months.