Signals Inbox·July 28, 2026·Frontier AI

Is Meta losing the AI race?

Meta has fallen behind the frontier-model leaders, but its advertising engine, recommendation systems and early lead in AI glasses make the broader race much harder to call.

Q1What would “losing the AI race” actually mean for Meta?

Meta is currently losing the race to build the world’s best general-purpose AI model, but that covers only one part of the contest that affects its future.

For Meta, “winning AI” has to mean more than finishing first on a benchmark. The company needs models good enough to avoid relying on rivals, an assistant people use deliberately, AI systems that keep improving its advertising business, and a credible position in whatever device comes after the smartphone.

The picture changes depending on where we look. Meta’s best current model remains about ten points behind the leader on Artificial Analysis. Its enterprise AI business is also tiny beside Anthropic, OpenAI and Google. At the same time, Meta’s recommendation and advertising models work across apps used by 3.56 billion people each day, and its partnership with EssilorLuxottica has already sold millions of AI glasses.

A model leaderboard captures only part of Meta’s position. The company can trail on difficult reasoning while leading in the systems that decide what billions of people watch, which ads they see and how they use an AI device worn on their face.

Today, the weaker models do not outweigh all those advantages. The gap is still too wide for Meta to qualify as an overall AI leader, though.

Meta’s position across the AI race

Q2Why did Meta suddenly look behind in AI?

Meta started to look behind because Llama 4 disappointed, its biggest promised model never arrived, and the company reacted with an unusually expensive reorganization.

Before that setback, Meta had a clear role in the market. OpenAI and Anthropic pushed proprietary models forward, while Meta made capable open-weight models widely available. Llama passed one billion cumulative downloads in early 2025, giving Meta enormous influence among developers even when it did not lead every benchmark.

Llama 4 weakened that position. Scout and Maverick arrived with useful features, including very long context, but neither became the model developers pointed to as the new open frontier. Behemoth created a bigger problem. Meta had presented it as the most capable member of the family and as the teacher behind the smaller models, then kept delaying its public release while competitors moved through several model generations.

Meta’s response showed how seriously it took the setback. It invested roughly $14.3 billion for a 49% stake in Scale AI, brought Scale founder Alexandr Wang into Meta, formed Meta Superintelligence Labs and recruited researchers from rival companies with exceptionally large pay packages. Meta later said the new group rebuilt its AI stack from the ground up over nine months.

This went far beyond a normal reshuffle. One of the largest investments in Meta’s history came with new leadership and a different model strategy. Meta plainly believed its old approach had stopped working.

Now the rebuilt organization is producing much better models. Meta has moved past defending Llama 4 and is trying to prove that the failure was temporary.

Q3How badly did Llama 4 hurt Meta?

Llama 4 badly hurt Meta’s credibility with developers because it ended the assumption that every new Llama generation would stay close to the best open models.

The old Llama strategy worked even when Meta ranked below the absolute leader. Developers could accept a modest quality gap in exchange for control, local deployment, customization and lower dependence on one API provider. Meta only needed to remain close enough for that trade-off to feel sensible.

Artificial Analysis’s current benchmark puts Llama 4 Scout near the bottom, far behind both Meta’s newer proprietary model and the frontier leaders. Its very long context window remains useful, but it cannot explain away a capability gap of that size.

Behemoth made the episode harder to dismiss. Delays happen with ambitious models, but Meta had already used Behemoth to support the story around the smaller Llama 4 releases. It then failed to deliver the flagship people were waiting for. By the time Muse Spark arrived, few observers still treated Behemoth as an imminent answer.

The timing also hurt. Developers were shifting from basic text generation toward coding, tool use, research and long-running agents. A broadly capable open model became less attractive when it failed too often during a multi-step task.

Llama still has a large installed base, and companies that already adapted it to their data will not switch casually. Even so, Llama 4 left Meta defending an ecosystem built by earlier models instead of setting the direction of open AI.

Q4Has Muse Spark 1.1 brought Meta back?

Muse Spark 1.1 has made Meta technically relevant again, although it still sits a full tier below the current leaders.

Artificial Analysis gives the model a score of 51, up eight points from the first Muse Spark in roughly three months. That places Meta alongside GLM-5.2 and below Grok 4.5, GPT-5.6 Sol, Claude Fable 5 and Claude Opus 5. The ten-point gap with the leader is meaningful, especially on demanding professional and agentic work.

The speed of improvement deserves attention. Muse Spark 1.1 raised its coding index from 59 to 71, reached 58% on SciCode and improved its professional knowledge-work score by 232 Elo. Its context window grew from 262,000 tokens to one million. Artificial Analysis also estimated a cost of about $0.26 per benchmark task, far below the roughly $2.03 measured for Claude Opus 5 at maximum effort.

One result needs more caution. The model’s hallucination rate fell from 73% to 38%, but its underlying accuracy stayed roughly flat while it refused more questions. For a real product, refusing an uncertain question is still better than inventing an answer. It tells us less about reasoning progress than the coding gains do.

Meta has also moved quickly from benchmark to product. Muse Spark 1.1 now powers Meta AI features that can connect to email and calendars, prepare slides, conduct web research, create recurring briefings and continue a task after the first answer. These features are still rolling out in select markets, so they show where the product is going rather than how widely people use it.

Meta has repaired the worst of the Llama 4 damage. Catching the leaders will require another release of similar quality, delivered before the frontier moves much further away.

Muse Spark 1.1 against selected frontier models

Q5Is Meta still behind in coding and AI agents?

Meta is still behind in coding products and reliable AI agents, even though Muse Spark 1.1 is now technically good enough to compete on several coding tasks.

The model reached 58% on SciCode, one of the strongest results Artificial Analysis had measured when it tested the release. Its broader coding index jumped by 12 points. Meta’s model team can now produce useful technical capability rather than another general chatbot with weak coding performance.

The product gap is much larger. Anthropic said Claude Code reached a $1 billion annualized revenue rate within six months of public launch and later reported a run rate above $2.5 billion. OpenAI and Google have also built large developer ecosystems around Codex, Gemini APIs and agent-development tools. Meta has disclosed no comparable coding product, revenue figure or developer habit.

Meta’s latest update gives the assistant real agent features. It can make plans, connect to email and calendars, generate presentations, schedule recurring briefings and change course while working. That is more useful than another demo showing a chatbot calling one tool.

What we still lack is evidence of reliability under pressure. A consumer may tolerate an imperfect restaurant plan. A software team will quickly abandon an agent that breaks a repository, misses permissions or needs constant correction during a long task. The strongest competitors have already spent months collecting feedback from those demanding workflows.

Meta can close part of this gap by placing agents inside WhatsApp, Instagram and its glasses. That would create uses competitors cannot copy as easily. For coding and professional agents, though, Meta is chasing products that already have users, revenue and established routines.

Q6Do people actually use Meta AI, or do they just see it inside Meta’s apps?

Meta AI has extraordinary reach today, but Meta has still not shown that people choose it as deliberately or use it as frequently as ChatGPT.

Meta’s last disclosed figure was more than one billion monthly users, largely because the assistant appears inside WhatsApp, Instagram, Facebook and Messenger. People can try it without downloading a new app, creating another account or learning an unfamiliar product.

ChatGPT shows what Meta has not told us. OpenAI reports more than 900 million weekly users, while Meta’s public milestone is monthly. A weekly figure points to much more frequent use than the same number measured over a whole month.

Google provides another useful baseline. Gemini has more than 900 million monthly users, up from 400 million a year earlier, and Google says daily requests grew more than sevenfold during that period. Meta has not published an equivalent growth rate for Meta AI requests, weekly users or daily-to-monthly engagement.

Meta said daily users generating media inside Meta AI tripled during the final quarter of 2025. Lately, Muse Image, deeper research, slide creation, recurring tasks and app connections have given people more reasons to return after one casual question.

We still need retention, prompt frequency and standalone-app usage before calling Meta AI a destination. Meta has made the assistant impossible to miss. It has not shown how many people come back on purpose or would miss it if the button disappeared.

Q7Is Meta already winning the AI advertising race?

Meta is already one of the clearest winners in applied AI advertising, and this part of the race is producing billions of dollars rather than leaderboard points.

The latest reported quarter gives us the clearest scale: $56.31 billion in revenue, up 33% year over year. Ad impressions rose 19%, the average price per ad increased 12% and operating margin held at 41%. Currency helped the reported growth, but revenue still rose 29% on a constant-currency basis.

Meta has also published results showing what these systems changed for advertisers. After doubling the GPUs used to train its Generative Ads Recommendation Model, the company recorded a 3.5% lift in Facebook ad clicks and more than a 1% increase in Instagram conversions. A separate runtime model increased conversions by 3% across Instagram Feed, Stories and Reels.

The creative side has reached a meaningful scale as well. Campaigns using Meta’s generative video tools were associated with a combined $10 billion annual revenue run rate in late 2025, and that activity was growing nearly three times faster quarter over quarter than overall advertising revenue. Meta’s incremental-attribution product reached a multi-billion-dollar run rate within seven months after a model update lifted measured incremental conversions by 24% compared with its standard attribution system.

Those figures need one clarification. Meta is measuring advertising revenue connected with the tools, rather than collecting $10 billion in software fees from an AI video product. Even so, advertisers are using the tools at a scale few standalone AI companies can reach.

Meta’s advantage comes from stacking small improvements across the whole advertising loop: creating the ad, selecting the audience, ranking the impression, predicting the purchase and proving that the purchase was incremental. A few percentage points at several stages can add up to a very large business result.

Q8Is recommendation AI Meta’s real advantage?

Recommendation AI remains Meta’s strongest technical advantage because the company has spent more than a decade training systems on billions of daily choices and can measure the outcome almost instantly.

A frontier chatbot must answer almost any question. Meta’s ranking systems solve narrower problems, but they do so at enormous speed and volume: which Reel to show next, which post belongs near the top of a feed and which ad is most likely to create a purchase.

The recent gains are large enough to see in user behavior. Meta reported that feed and video-ranking improvements raised views of organic Facebook posts by 7%. Threads optimizations increased time spent by 20%. Instagram lifted the share of recommended content coming from original posts to 75% in the United States, while hundreds of millions of people watch AI-translated videos each day.

The gains feed into one another. More relevant content keeps people watching and creates more ad opportunities; stronger ranking and measurement make advertisers willing to pay more for them. In the latest quarter, impressions and price both rose by double digits, a much stronger result than growing one by sacrificing the other.

Meta’s social data does not automatically give it the best general reasoning model. Much of that data is private, noisy or poorly suited to advanced coding and scientific tasks. It is extremely useful for understanding taste, relationships, creators, products and behavior.

That explains the split. Meta is behind when a model must solve an unfamiliar expert problem. It is often ahead when AI has to predict what a specific person will watch, click or buy next. For Meta’s existing business, the second skill makes more money. Frankly, it is not close.

Q9Has Meta lost the open-source AI lead?

Meta has lost the open-weight AI lead for now, even though Llama remains one of the most widely used model families in the world.

Llama crossed one billion cumulative downloads in early 2025. That gave Meta a real ecosystem: cloud deployments, fine-tunes, local versions, research projects and companies that built internal tools around the models. Those installations do not vanish because a newer benchmark looks weak.

Leadership has still moved elsewhere. Chinese developers including DeepSeek, Alibaba, Zhipu and Moonshot have released strong open or open-weight models across reasoning and coding. Artificial Analysis currently scores GLM-5.2 at 51, level with Muse Spark 1.1 and far above Llama 4 Scout. Google’s Gemma family also gives developers another well-supported option.

The cumulative Llama number combines generations, sizes, repeated downloads and different versions. It proves that Llama reached a huge audience, while telling us little about how many developers are starting new projects with the latest release.

Meta’s own recovery model makes the shift especially clear. Muse Spark 1.1 is proprietary and available through Meta’s API. Meta has said future open releases remain possible, but the model currently repairing its technical reputation cannot be downloaded and run independently.

Keeping Muse Spark proprietary may help Meta charge for usage and control the product, but it weakens the old promise behind Llama. Meta now has to compete for API users against companies with stronger enterprise relationships, while open-model developers can choose alternatives whose weights are available today.

Llama still gives Meta influence, familiarity and a large installed base. These days, it looks more like a huge installed platform than the model family setting the open frontier.

Q10Did Meta’s AI reorganization actually work?

Meta’s AI reorganization has produced a convincing first result, but one fast model cycle is too little evidence to declare the organization fixed.

The reset addressed obvious problems. The Scale AI deal brought Meta closer to a major provider of training data and evaluations. Alexandr Wang became the operational leader of Meta Superintelligence Labs. New hires joined existing researchers across training, post-training, product and infrastructure.

Muse Spark is the strongest evidence that the new setup can ship. Meta said the group rebuilt its model stack over nine months, released its first model and followed with an eight-point benchmark improvement roughly three months later. A shipped model with measurable gains tells us far more than another list of expensive hires.

The products are also arriving faster. Muse Spark moved into Meta AI, Muse Image followed, and Meta opened a first-party API for Muse Spark 1.1. The releases now build on one another instead of appearing as isolated research announcements.

A nine-month rebuild proves speed, not stability. Meta assembled much of the organization quickly, and another leadership change would reopen the same problem. The Scale deal also made a previously neutral supplier look closely tied to one customer, prompting some rival labs to reduce their exposure.

We will know the reorganization worked when Meta delivers several strong models on schedule and keeps the core team together. For now, the reset looks promising. It still needs another strong model cycle.

Q11Can Meta afford this AI spending spree?

Meta can afford its current AI spending spree, although the size of the increase leaves little room for years of weak returns.

The company generated $200.97 billion in revenue and $83.28 billion in operating income during 2025. Even after $72.22 billion of capital expenditure, free cash flow reached $43.59 billion. It entered 2026 with roughly $81 billion in cash, equivalents and marketable securities.

The pressure comes from how much faster capital spending is rising than the business. Meta expects $125 billion to $145 billion in 2026 capital expenditure. The midpoint, $135 billion, is about 87% above the 2025 total and equal to roughly two-thirds of last year’s revenue.

The core business is still growing fast enough to carry the bill. Meta’s latest quarter produced 33% revenue growth, $22.87 billion in operating income and $12.39 billion in free cash flow after nearly $20 billion of capital expenditure. Management still expects annual operating income to exceed its 2025 result.

The urgency is obvious. Spending jumped after the Llama setback, the Scale investment and the creation of Meta Superintelligence Labs. Meta is buying room for repeated attempts because falling behind the next computing platform could be far more costly than a few years of excessive infrastructure.

The plan also has more discipline than a simple GPU shopping spree. Meta can use the same capacity for recommendation systems, ad ranking, content tools, Meta AI and glasses. Its custom chips handle predictable inference workloads, while outside suppliers cover the most demanding training jobs.

The risk will show up later in depreciation and operating costs. Data centers keep hitting the income statement long after construction ends. If Meta AI remains lightly used or model efficiency improves faster than demand, today’s investment could produce mediocre returns for years.

Meta’s financial capacity for AI investment

Q12Can Meta reduce its dependence on Nvidia and other chipmakers?

Meta can reduce that dependence for everyday AI workloads, but it still needs outside chips for the largest training runs.

The company says it has deployed hundreds of thousands of MTIA chips across organic recommendations and advertising. MTIA 300 is already in production for ranking and recommendation training. Meta plans three more generations, with MTIA 450 and 500 designed mainly for generative-AI inference.

That focus fits Meta’s economics. Training a frontier model is extremely expensive, but serving the model to billions of possible users can cost even more over time. A small saving on each request becomes important when the assistant appears across WhatsApp, Instagram, Facebook, Messenger and glasses.

Meta is also moving faster than normal chip cycles. It plans four generations within two years and says its modular design can support releases every six months or less. The chips use common systems such as PyTorch, vLLM and Triton, which should make them easier for Meta’s engineers to adopt.

Meta still buys from Nvidia, AMD and other suppliers, while working with Broadcom, Arm and AWS on parts of its infrastructure. Different jobs need different chips. Google has a longer record of training frontier models on its own TPUs, so Meta remains more dependent on outside hardware.

The goal is straightforward: cut the cost of recurring workloads and save expensive general-purpose chips for the hardest jobs.

Q13Could AI glasses become Meta’s biggest advantage?

AI glasses currently give Meta its best chance to lead a major new AI category from the start.

When we look at actual sales, the lead is already substantial: EssilorLuxottica reported more than seven million AI glasses sold during 2025 across Ray-Ban Meta and Oakley Meta products. Sales accelerated during the year, with the company saying second-half growth was much faster than first-half growth. These are paid consumer devices sold through normal eyewear channels, which gives the result more weight than a prototype demonstration.

Glasses also give AI a reason to exist outside a chat box. The assistant can see what the wearer sees, hear a question, translate speech, capture media and answer without forcing the user to take out a phone. Muse Spark’s multimodal abilities and Meta’s latest agent features fit that interface naturally.

The partnership solves a problem technology companies often underestimate. EssilorLuxottica supplies recognizable brands, lenses, design, manufacturing and retail distribution. Meta supplies the assistant, software, apps and cloud infrastructure. Together, they have produced a device people are willing to wear in public.

The category remains small beside smartphones and ordinary eyewear. Battery life, privacy, social acceptance and limited displays constrain what current products can do, while Apple and Google have the hardware ecosystems to enter more aggressively.

Meta’s current lead is still real because the product has crossed from experiment to a genuine consumer category. If glasses become a common way to use AI, Meta will own a valuable interface while rivals remain more dependent on Apple’s and Google’s phones.

Q14Why is Meta so weak in enterprise AI?

Meta is currently far behind in enterprise AI because it lacks the cloud platform, sales channel and trusted workplace product that turn a strong model into a large business.

Anthropic says its annualized revenue has passed $30 billion, up from roughly $9 billion at the end of 2025. More than 1,000 customers now spend over $1 million each on an annualized basis, and Claude Code alone has surpassed a $2.5 billion run rate. OpenAI reports more than nine million paying business users. Google distributes Gemini through Workspace, Cloud and its enterprise agent platform.

Meta has disclosed no similar figure for Muse Spark, its API or workplace assistants. The company shut down Workplace, and it does not operate a general cloud platform comparable with AWS, Azure or Google Cloud. A bank or software company buying AI usually wants security, databases, identity tools, deployment controls and support in the same contract. Meta cannot offer that full stack today.

Its business-messaging position is stronger. Paid WhatsApp messaging crossed a $2 billion annual run rate in late 2025, and early business AI trials in Mexico and the Philippines were already handling more than one million weekly conversations. Meta can become important for customer support, product discovery and commerce inside messaging.

That opportunity is much narrower than enterprise AI as a whole. Helping a shop answer questions on WhatsApp involves different budgets and requirements from running coding agents or regulated internal workflows across a multinational company.

Those larger customers also provide demanding feedback and recurring payments. Meta can finance research through ads, but it gets less direct evidence about whether outside companies consider its general-purpose AI essential.

Q15What would prove Meta has caught up in AI?

Meta will have caught up when strong models, frequent user behavior and outside commercial adoption appear together, rather than one at a time.

First, Meta needs another model release near the frontier. Two strong generations in a row would show that the Muse Spark improvement can be repeated.

Second, Meta should publish weekly usage, retention and prompts per user for Meta AI. The billion-user monthly figure shows how many people can reach it; a strong weekly-to-monthly ratio would show that they return.

Third, the API needs outside demand. We would expect Meta to disclose active developers, token volume, enterprise customers or direct revenue if Muse Spark becomes a serious platform. Without those figures, most of the model’s success remains inside Meta’s own products.

Fourth, the spending has to pass through the income statement without crushing profit growth. Meta plans to spend around $135 billion at the midpoint of its capital-expenditure guidance. Continued operating-income growth after depreciation rises would show that AI is paying for more than its electricity and hardware.

Finally, Meta needs one product competitors cannot easily neutralize. Glasses are the leading candidate. Tens of millions of active wearers using Meta AI daily would give the company control over a new interface and make small benchmark gaps less decisive.

Q16So, is Meta losing the AI race?

Meta is losing important parts of the AI race today, but calling the whole company an AI loser is no longer supported by the evidence.

The company remains clearly behind in frontier intelligence, with Muse Spark 1.1 still about ten benchmark points below the leader. Meta also trails badly in enterprise revenue, coding products and the evidence that users deliberately choose its assistant.

Its earlier failure was serious. Llama 4 fell far below expectations, Behemoth never became the promised flagship and Meta responded with a roughly $14.3 billion Scale AI investment, a new lab and one of the fastest capital-spending expansions in corporate history.

The recovery is now substantial enough to change the answer. Muse Spark improved by eight benchmark points in about three months and is already cheap enough to serve consumer products at a much lower measured cost than the most capable Claude configuration. Frontier performance remains out of reach for now.

Meta also leads where the race touches its current business. AI is improving feed engagement, ad clicks, conversions, creative tools and measurement across a platform used by 3.56 billion people daily. The latest quarter combined 19% growth in ad impressions with a 12% rise in price per ad while maintaining a 41% operating margin. AI glasses have already crossed seven million annual unit sales, giving Meta an early lead in a category that could matter far more than another chatbot app.

The verdict: Meta is behind, recovering quickly and still dangerous. It has lost the right to call itself a frontier-model leader, and it may never become the preferred enterprise AI provider. Yet its distribution, advertising engine, infrastructure and glasses business give it several ways to win without owning the single smartest model.

The answer could turn more negative if the next Muse release stalls or if Meta AI’s enormous monthly reach never becomes frequent use. For now, Meta is losing the most visible AI contest while remaining unusually strong in the contests that could produce the largest consumer reach and the clearest profits.