OpenAI and Anthropic Built a Lead. xAI and Meta Want to Blow It Up.

For a while there, the AI race was starting to look almost simple.

OpenAI had ChatGPT, Codex, a massive consumer audience, and models that reliably sat somewhere around the top of the frontier. Anthropic had Claude, Claude Code, enormous enterprise momentum, and a reputation for building AI that was particularly good at coding and actually getting work done.

Everyone else was chasing.

That doesn’t mean OpenAI and Anthropic had AI sewn up. Google was never exactly some scrappy also-ran, and Meta, xAI, DeepSeek, Alibaba, Moonshot AI and a growing list of other labs weren’t going anywhere.

But OpenAI and Anthropic had built something more important than a temporary benchmark lead. They had turned frontier intelligence into products people were actually using.

By February, Anthropic said Claude Code alone had passed a $2.5 billion annualized revenue run rate, more than doubling since the beginning of the year. Business subscriptions had quadrupled, and enterprise customers accounted for more than half of Claude Code revenue.

OpenAI, meanwhile, has spent 2026 pushing ChatGPT, Codex and its growing collection of agentic tools toward something much broader than a chatbot. Its current GPT-5.6 family is explicitly built around everything from high-volume routine work to long-running knowledge work, coding and scientific problems. OpenAI also cut prices sharply in July, dropping GPT-5.6 Luna pricing by 80% and Terra by 20%.

The two companies have been trying to turn their models and tools into the places where people and companies actually do work.

And then xAI came screaming back into the race.

Elon Musk has never lacked ambition in AI. For much of the past year, though, the question was whether xAI could turn enormous amounts of money, GPUs and bravado into models that consistently competed with the very best systems from OpenAI and Anthropic.

Grok 4.6 makes that question considerably harder to ask with a straight face.

Independent testing from Artificial Analysis gives Grok 4.6 a score of 61 on its Intelligence Index, a composite measure spanning reasoning, knowledge, math and coding. That puts it alongside OpenAI’s GPT-5.6 Sol and behind only Anthropic’s leading systems among the models Artificial Analysis has tested.

The number by itself is impressive. The trajectory, however, seems to tell us more.

Grok 4.6 gained five points over Grok 4.5 after only about a month. Compared with Grok 4.3, xAI has gained 23 points on the same index. Artificial Analysis now describes the company as having returned to the intelligence frontier.

That is a hell of a rate of improvement.

It’s also happening while xAI is undercutting some of its biggest competitors on price.

Grok 4.6 costs $2 per million input tokens and $6 per million output tokens through the xAI API. Artificial Analysis calculates its cost at about 84 cents per Intelligence Index task, giving it one of the strongest cost-to-intelligence ratios among frontier models.

That changes the competitive equation.

xAI doesn’t necessarily have to produce a model that is indisputably smarter than Claude or GPT on every benchmark. If it can stay within striking distance while selling that intelligence significantly cheaper, it can put pressure on the entire frontier market.

We’ve already seen how quickly that kind of pressure spreads. OpenAI’s July price cuts are a useful reminder that model pricing is hardly carved into stone.

There’s also a tendency in AI coverage to treat benchmark rankings like Formula 1 standings. One model moves two spots up the table, another drops three, and suddenly we’re declaring a new world champion.

I think the more interesting question here is whether the gap itself is becoming easier to close.

If the answer is yes, OpenAI and Anthropic have a much more complicated problem than Grok 4.6 having a good week. Because here’s what Musk says to expect next…

xAI built Grok 4.6 it specifically with long-running agents in mind.

The company says Grok 4.6 is designed to stay with complex tasks across many steps, including research, analysis, coding and building finished applications or work products.

And just before releasing it, xAI introduced Grok Bot.

Grok Bot is considerably more ambitious than putting a chatbot in another sidebar.

Each Bot gets its own persistent computer in the cloud, complete with a browser, filesystem and terminal. It can sign into tools, work inside applications and websites, use the command line, and keep running even when the user’s laptop is closed. Multiple Bots can work in parallel.

You give one work. It goes and tries to finish it. That puts xAI directly into the rapidly emerging fight over autonomous AI labor.

Anthropic has Claude Code and has been pushing Claude deeper into long-running agentic work. OpenAI has Codex and is steadily expanding its own agent infrastructure. Microsoft is building agents throughout its software stack. Everyone appears to have arrived at roughly the same conclusion: chat was the opening act.

The valuable AI systems of the next few years may be the ones that can be trusted to go away for an hour, operate software, coordinate multiple steps, make reasonable decisions and come back with the job actually done.

Grok Bot is xAI planting a flag squarely in that territory.

It’s also notable that xAI says its own teams already used an internal version of the system for sales, marketing, operations and software work before releasing the product publicly. The current beta is being positioned as an AI teammate rather than another assistant.

Grok 4.6 gets xAI into the frontier-model conversation. Grok Bot is the attempt to turn that technical credibility into a business people actually build around.

And that’s where xAI starts looking significantly more dangerous labs that have been firmly at the top of the leaderboard for years.

There is, of course, another ingredient in xAI’s rise: an absolutely ridiculous quantity of hardware.

The company’s Colossus buildout in Memphis has become one of the defining infrastructure projects of the AI boom. And xAI’s ability to bring enormous amounts of compute online quickly is becoming a strategic asset of its own.

The weirdest evidence for that might be the identity of one of its biggest customers: Anthropic.

Earlier this year, Anthropic struck a deal to use the full computing power of SpaceX’s Colossus 1 facility, which contains more than 220,000 Nvidia processors and supplies roughly 300 megawatts of capacity.

Think about that for a second.

One of the companies xAI is trying to catch is also paying for access to the infrastructure helping make that chase possible.

SpaceX has since been pursuing AI compute as a business in its own right, including additional deals with Google. Reuters reported in June that the Google agreement covers access to roughly 110,000 Nvidia GPUs, while Musk has publicly discussed offering AI compute as a service at significant scale.

There’s an almost comical amount of vertical integration beginning to emerge here. For example, the new massive Terafab facility is expected to be the largest building ever constructed on earth.

Models. Agents. Coding tools. Data centers. Compute services. SpaceX infrastructure. X distribution.

You can argue about how well all of those pieces fit together, and there are plenty of reasons to question whether xAI can turn them into a coherent enterprise business. Earlier this year, Grok still lagged badly behind OpenAI and Anthropic in enterprise adoption. Building a great model doesn’t magically create trust, developer loyalty or a mature ecosystem.

But the fundamental question around xAI has changed. A year ago, it was reasonable to ask whether Musk could spend his way into the frontier.

Now xAI is there. Can stay there long enough to turn that technical progress into something much larger?

And if xAI represents the most direct attack on OpenAI and Anthropic’s model lead, Meta is coming at the problem from almost the opposite direction.

If xAI represents the most direct attack on OpenAI and Anthropic’s lead, Meta is coming at the problem from almost the opposite direction.

For much of the past year, Meta’s AI story was easy to summarize and not especially flattering.

Llama 4 landed with considerably less impact than Meta hoped, to put it gently. The company reorganized its AI operation, spent aggressively to recruit talent, assembled Meta Superintelligence Labs, and started pouring extraordinary amounts of money into infrastructure.

The implied message was pretty clear: Meta knew it had fallen behind.

That’s becoming harder to say now.

Meta has released three generations of Muse Spark in four months, and the newest, Muse Spark 1.2, scores 57 on the Artificial Analysis Intelligence Index. That puts Meta firmly back among the cluster of companies producing models close to the frontier. Artificial Analysis says the latest improvement was particularly strong on agentic evaluations, the kinds of tests meant to measure whether a model can actually complete useful work across multiple steps.

Meta also released Muse Code, a coding agent powered by Spark 1.2 that can work through long software-development tasks and run multiple sub-agents in parallel. Meta is charging developers $1.25 per million input tokens and $4.25 per million output tokens, putting it directly into the same developer market OpenAI, Anthropic and now xAI are fighting over.

So Meta suddenly has a much more conventional competitive story than it did a few months ago.

It has a good model, an agentic coding product, an API, and it has enough money to spend as much as $145 billion on infrastructure this year alone.

But I don’t think any of those are the most interesting parts of Meta’s strategy.

Because Meta also appears increasingly interested in changing the rules of the game.

On August 10, Mark Zuckerberg published a sprawling essay called “The Future is for Everyone.”

The title sounds a bit like something printed on a poster in an elementary-school guidance counselor’s office, but the argument underneath it is consequential.

Zuckerberg lays out a vision in which highly capable personal agents become widely available and deeply integrated into people’s lives. These agents would know your goals, interests and preferences, and work continuously on your behalf across everything from careers and hobbies to relationships and home management.

Meta also used the occasion to reembrace something it once looked ready to abandon: open-weight AI.

The company released Muse Glimmer, a 30-billion-parameter model designed to run agentic tasks locally on a Mac or PC with a single GPU, and announced that it plans to release the weights for the much more powerful Muse Spark 1.2 as well.

That matters because Meta has an economic incentive that looks very different from OpenAI’s or Anthropic’s.

OpenAI needs people and companies to pay for access to its models and products. Anthropic needs people and companies to pay for Claude. Meta already owns Facebook, Instagram, WhatsApp and Messenger. Its family of apps reaches about 3.6 billion people every day.

Meta can make money when AI makes those products more useful, keeps people engaged, improves advertising, sells hardware, powers commerce, or creates entirely new experiences.

It doesn’t necessarily need the model itself to be the product, which opens up an intriguing, and rather unique, strategy.

Meta can afford to make increasingly capable AI cheap, open or ubiquitous because it has enormous businesses waiting downstream.

And that could become deeply uncomfortable for companies whose business models depend more directly on charging for intelligence.

If Meta releases a genuinely competitive open-weight model that companies can download, customize and run themselves, it doesn’t have to steal every Claude or ChatGPT customer to have an impact.

It can push the market price of intelligence downward. It can make proprietary APIs harder to differentiate. It can encourage developers to build around models they control rather than models rented from someone else.

In other words, Meta may be able to hurt the economics of the frontier-model business simply by being Meta.

There is a certain historical symmetry to that.

Meta spent years building enormous businesses on top of technologies it did not have to invent or own outright. Now it may be in a position to make the underlying AI layer more abundant and build its next generation of products on top of that.

And unlike almost every AI startup in existence, Meta doesn’t have a distribution problem.

It has the opposite problem. It has so many places to put AI that figuring out which ones people actually want may be the harder question.

Imagine building a new AI company and being told you could choose your launch strategy.

Would you like to start with an empty website and buy ads? Or would you prefer immediate access to Facebook, Instagram, WhatsApp, Messenger, smart glasses and billions of existing users?

Meta gets option two.

Muse Spark is already being integrated into Meta AI, and the company has said its newer models will increasingly replace the Llama systems previously powering AI features across WhatsApp, Instagram, Facebook and its smart glasses.

That changes what adoption means.

OpenAI had to convince hundreds of millions of people to seek out ChatGPT.

Meta can put its AI inside products people already open dozens of times a day.

And its hardware ambitions make that distribution strategy even more interesting.

If the next major interface for AI is some combination of phones, glasses, earbuds and ambient agents rather than a browser tab, Meta is one of the few companies already shipping consumer hardware specifically designed around that future.

This is where the “personal superintelligence” rhetoric starts to feel more like a product roadmap.

A model that knows you well, sees some of what you see, communicates through products where your friends and family already live, and can act on your behalf is a very different proposition from opening a chatbot and typing a prompt.

It also brings us to Meta’s biggest problem. That system would have to know an awful lot about you.

There are at least two reasons to be skeptical of the Meta comeback story.

The first is simple: it has to keep delivering.

Muse Spark 1.2 is a substantial improvement, and at 57 on Artificial Analysis it belongs in the same conversation as several frontier systems. But Meta has had promising AI moments before. Llama 4 is a useful reminder that an enormous research organization and unlimited resources don’t guarantee a smooth march toward the frontier.

The interesting signal is the velocity. Muse Spark arrived in April. Spark 1.1 followed in July. Spark 1.2 arrived in August.

Artificial Analysis measured Spark 1.1 at 51 on the Intelligence Index and Spark 1.2 at 57 at its highest reasoning setting. That is meaningful progress in a very short period.

If Meta maintains that pace, the argument that it is fundamentally behind becomes difficult to sustain.

The second problem may be harder.

Meta’s vision requires trust on an almost comical scale. A truly useful personal agent needs access to your messages, contacts, interests, schedule, location, purchases, social relationships and potentially what you’re seeing through wearable devices.

Meta would very much like to be the company holding that context. There will, however, be people who hear that sentence and immediately begin looking for the emergency exit.

Meta doesn’t need everyone to trust it. Three-and-a-half-billion daily users provide a fair amount of margin for error.

But if deeply personal AI becomes one of the major computing platforms of the next decade, trust becomes a competitive feature, and Meta carries more historical baggage there than some of its rivals.

Still, the strategic position is hard to dismiss.

Meta now has increasingly competitive models, a developer API, coding agents, open-weight ambitions, consumer hardware, staggering amounts of compute and one of the largest distribution networks ever assembled.

Which raises an awkward question about the premise of this entire article.

At this point, someone at Google is probably yelling at the screen.

Fair enough.

Any argument that begins with OpenAI and Anthropic as AI’s “top two” immediately runs into a fairly enormous Gemini-shaped objection.

Google is not a distant challenger trying to break into AI. Google has DeepMind. Google has Gemini. Google has its own TPUs. Google has Google Cloud. Google has Search, Android, Chrome, Workspace and YouTube.

And Google has the kind of distribution numbers that make even Meta’s look merely ridiculous rather than completely absurd.

As of Google’s latest earnings update, the Gemini app has 950 million monthly active users, with daily users tripling over the previous year.

AI Mode in Search has already passed 1 billion monthly users, while AI Overviews reach more than 2.5 billion people each month. Those are civilization-scale distribution numbers.

So why frame this around OpenAI and Anthropic in the first place?

Because “who leads AI?” turns out to depend heavily on what exactly you mean by leading.

If we’re talking about the companies that have most clearly defined the recent frontier around general-purpose models, coding systems and increasingly autonomous agents, OpenAI and Anthropic have an excellent claim.

If we’re asking who has the strongest overall position in AI, the answer becomes much less obvious.

Google may have the most complete stack of anyone in the race. It designs AI chips. It owns data centers. It trains frontier models. It sells those models through Google Cloud. It distributes AI through Search. It embeds Gemini into productivity software. It controls the operating system on most of the world’s smartphones. It owns the world’s largest video platform.

And unlike companies spending billions to figure out how to monetize AI someday, Google already has several enormous businesses capable of absorbing and monetizing it.

The problem for Google has often been less about possessing the pieces than making all of them feel like one coherent AI strategy. Which means Google doesn’t really belong in a section about companies trying to catch OpenAI and Anthropic, but their lack of a frontier competitive model for nearing a year makes it a relevant question.

Google is the enormous incumbent hiding in plain sight. And Google isn’t even the biggest problem with treating this as a four- or five-company race.

Pull up the Artificial Analysis Intelligence Index and something else becomes immediately obvious. There are a lot more logos near the top than there used to be.

As of mid-August, Moonshot AI’s Kimi K3 scores 57 on the Intelligence Index. Z AI’s GLM-5.2 sits just behind the leading group, while DeepSeek V4 Pro scores 53. Alibaba’s Qwen family also remains among the strongest model lineups coming out of China.

Artificial Analysis now maintains an entire open-model leaderboard where Kimi, Alibaba, DeepSeek and Z AI occupy several of the strongest positions.

A few years ago, the frontier of generative AI was overwhelmingly discussed as a competition between a small collection of American companies.

Now some of the strongest systems in independent testing are being developed by Chinese labs, and several of them are available with open weights.

Moonshot’s Kimi K3 is particularly interesting.

Artificial Analysis found that Kimi K3 performs at a level comparable to leading proprietary systems on its broader Intelligence Index, while also ranking near the very top on its AA-Briefcase benchmark for agentic knowledge work.

DeepSeek, meanwhile, continues to push hard on economics. Its latest V4 Pro scores 53 while costing Artificial Analysis about 25 cents per Intelligence Index task.

Cheap, capable open models can spread differently from proprietary systems. Companies can host them themselves. Researchers can modify them. Developers can fine-tune them without waiting for permission from a model provider. Entire ecosystems can form around them.

And improvements can propagate through the market in ways that make it increasingly difficult for any single company to maintain a durable technical advantage.

There is a geopolitical layer here, too.

The United States and China are increasingly treating advanced AI as a strategic competition between rival technological blocs.

The technology itself looks considerably messier.

The frontier is becoming more multipolar at almost exactly the moment governments are trying to make the geopolitical race more binary. That tension probably deserves its own article.

For this one, the more important point is simpler.

OpenAI and Anthropic didn’t suddenly get worse, their competition got much better.

And once you zoom out far enough to include xAI, Meta, Google, Moonshot, Alibaba, DeepSeek and Z AI, something starts to look increasingly strange:

We are still talking about the AI race as though it has a single leaderboard.

Benchmark rankings give us a clean answer: Model A scores higher than Model B, so Model A is winning. The industry is getting harder to measure that way.

Once several companies can build models within striking distance of the frontier, other advantages start to matter more: price, agents, enterprise adoption, developer ecosystems, distribution, compute and trust.

The AI race increasingly looks more like a decathlon.

OpenAI has consumer scale and a growing agent ecosystem. Anthropic has coding and enterprise momentum. xAI has rapidly improving models, aggressive pricing and enormous compute. Meta has openness and distribution. Google has perhaps the most vertically integrated stack of them all.

And Chinese labs are adding even more pressure through increasingly capable, often cheaper and more open models.

There may never be one company that wins every category. That may be the more useful way to think about the next phase of AI.

That is why I’m increasingly skeptical that the next phase of AI will look anything like the first.

For the first few years of generative AI, the industry raced to build intelligence.

Now it is racing to build everything around it.