Update: This analysis has been revised to add independent benchmark context, distinguish model performance from agent endurance, incorporate Meta's latest engagement and financial results, and more fully examine the risks of its capital spending.
Futurism's July 31 article argues that Meta has "almost nothing to show" for its enormous AI spending and is "practically absent from the frontier AI model race."
That conclusion is difficult to reconcile with the available evidence. Independent evaluations place Muse Spark 1.1 among competitive frontier models. Meta reports early use of its business and advertising tools, and the company can distribute AI products across applications serving 3.60 billion people each day. Its latest results also show growth in users, engagement, advertising and revenue rather than the continuing decline described by Futurism.
None of that proves Meta will earn an acceptable return on its infrastructure investments. The company expects $130 billion to $145 billion in full-year capital expenditures, while consumer demand for its vision of personalized AI remains uncertain. The more defensible assessment is not that Meta has produced almost nothing, but that it has produced credible technology and early commercial signals at a cost that creates an unusually high bar for success.
Evaluating that position requires separating three questions that Futurism largely combines: whether Meta's models are technically competitive, whether its distribution can generate durable adoption, and whether either advantage can justify the capital required.
Muse Spark is competitive, with identifiable weaknesses
Muse Spark 1.1 trails leading systems in some long-running agentic tasks. Meta has acknowledged those gaps, and Axios reported that agents from OpenAI, Anthropic and Google can currently operate for longer and handle a wider variety of jobs.
That is a meaningful limitation, particularly as developers increasingly evaluate models by their ability to complete extended workflows rather than answer isolated prompts. It is not, however, evidence that Muse Spark is broadly outperformed across most tasks.
Independent evaluator Artificial Analysis scored Muse Spark 1.1 at 51 on its Intelligence Index, effectively tying it with GPT-5.4 at xhigh reasoning, GPT-5.6 Luna and GLM-5.2. The model scored 45% on Humanity's Last Exam, one point behind Claude Opus 4.8, and ranked third among all tested models on SciCode.
Muse Spark also has a one-million-token context window. It completed Artificial Analysis's evaluation with fewer output tokens than the similarly scoring GPT-5.4, GPT-5.6 Luna and GLM-5.2. At Meta's API prices, Artificial Analysis estimated a cost of roughly 26 cents per Intelligence Index task.
Benchmarks are not a complete measure of product quality. They can favor particular model configurations, and strong scores do not establish reliability in production or adoption among developers. But they provide a more concrete basis for comparison than a general characterization that Meta is absent from the frontier. Artificial Analysis published the complete results on July 10, three weeks before Futurism's article.
The evidence supports a narrower conclusion: Muse Spark is an economically competitive frontier model with weaknesses in extended agent operation and no demonstrated developer position comparable to OpenAI or Anthropic.
Muse Image warrants scrutiny rather than dismissal
Futurism gives one sentence to Meta's image model:
"At the same time, Meta released an image-generation model called Muse Image, which felt like an afterthought and a too-little-too-late attempt to catch up with its competitors."
Meta reported at launch that Muse Image held the No. 2 position on Arena's human-preference rankings in three categories: text-to-image generation, single-image editing and multi-image editing. Its forthcoming Muse Video model ranked third for text-to-video.
The image model can search for visual references, write and execute code when a request requires precision, revise its output, preserve coherence across editing turns and combine people, objects, clothing and environments from multiple reference images. It also works with Muse Spark on tasks including website creation and animated media.
Those capabilities and rankings came through Meta's own announcement, so they should not be accepted without qualification. A rigorous assessment would examine Arena's methodology, test whether the model performs consistently outside benchmark conditions and determine whether its capabilities translate into regular use. Meta has not yet established that competitive image rankings will produce a durable consumer or developer business.
The technical announcement nevertheless contains enough detail to require analysis. Describing the product as an afterthought without engaging with its performance, architecture or intended integration does not resolve whether Meta is catching up or building a differentiated system.
Agent endurance is not the same as overall model performance
Futurism writes that Muse Spark is "still easily outdone in most tasks by competing models from OpenAI, Anthropic, and Google," attributing that assessment to Axios.
The linked Axios report makes a more specific point: rival agents can handle a broader range of tasks and operate autonomously for longer. That distinction matters. Agent endurance measures whether a system can plan, use tools and recover from errors across lengthy workflows. It is strategically important, but it is one dimension of performance rather than a proxy for every coding, reasoning, scientific or long-context task.
A separate Axios report about the Muse Spark 1.1 launch noted the model's improved coding and long-context abilities, competitive API pricing and deployment across Facebook, Instagram, WhatsApp and Meta AI.
Taken together, the reports describe a model with competitive underlying capabilities and a weaker position in extended autonomous work. Turning that weakness into a claim about "most tasks" goes beyond the cited reporting.
Distribution gives Meta an advantage, not a guaranteed audience
Meta averaged 3.60 billion daily users across its applications in June, up 3% from the previous year. Instagram has reached two billion daily users, Facebook has more than two billion, and Meta's Threads app has passed 500 million monthly users.
That reach changes the economics of product distribution. Meta can place AI features inside WhatsApp, Instagram, Facebook and Messenger rather than relying entirely on application downloads, account creation and new consumer habits. Meta AI chief Alexandr Wang told Axios that the billions of people using Meta's products, combined with the company's knowledge about those users, represent an advantage competitors cannot replicate.
Distribution should not be confused with demand. Meta AI remains less culturally prominent than ChatGPT, Claude or Gemini. Some users resent its placement in WhatsApp and Instagram, and Meta's history creates substantial privacy and trust concerns for an assistant designed to use personal context. A product can be widely available without becoming indispensable.
Meta's reported usage figures provide early but incomplete evidence. The company says daily interactions with its assistant have increased 60% since it rebuilt Meta AI around Muse Spark. More than one million businesses use Meta's business agents each week, while nine million small businesses use at least one of its AI advertising tools.
These are company-reported figures, and Meta has not disclosed enough underlying data to establish frequency, retention or the depth of consumer engagement. They do show that the company has more than model benchmarks to show for its spending. The unresolved issue is whether this activity reflects durable adoption or exposure created by Meta's control of distribution.
The latest quarter does not show continuing engagement declines
Futurism supports its claim that Meta's users and engagement "continue to slide" by linking to an April Verge article about a sequential decline of 20 million daily users in the first quarter. Meta attributed that decline to internet disruptions in Iran and WhatsApp restrictions in Russia.
By the time Futurism published, Meta had released its second-quarter results. Daily users had risen to 3.60 billion. Instagram time spent increased by double digits from the previous year, while Facebook video time grew 9% globally and by more than 10% in the United States and Canada.
Meta's ad impressions increased 14%, and the average price per advertisement rose 12%. Revenue grew 28% to $60.8 billion, while Family of Apps advertising revenue increased 27%.
Those figures do not settle questions about the quality of Meta's feeds or the long-term effect of synthetic content. They do contradict the narrower factual claim that user numbers and engagement were continuing to decline. A current assessment should account for the latest quarter rather than treating the previous sequential dip as the continuing trend.
The financial case remains unresolved
The strongest argument against Meta's AI strategy concerns returns, not whether the company has produced capable models.
Meta spent $31.1 billion on capital expenditures during the quarter, while free cash flow fell to $784 million. It now expects $130 billion to $145 billion in full-year capital expenditures and has signaled additional infrastructure commitments beyond that period.
This scale creates several risks. Muse Spark has limited developer adoption compared with OpenAI and Anthropic. Meta's personal-agent strategy remains largely aspirational. The company has not demonstrated that consumers want the "nearly infinite universe" of personalized AI content Mark Zuckerberg described during the earnings call. For users already frustrated by synthetic material in their feeds, that vision may reduce trust rather than deepen engagement.
Meta also recorded $2.4 billion in legal charges and $1.2 billion in severance expenses. Its workforce is undergoing repeated reorganizations while management commits the company to infrastructure that could produce years of depreciation and operating costs. Even if Meta's models remain competitive, the spending can destroy value if demand, pricing or utilization falls short of the capacity being built.
The central tension is therefore more consequential than Futurism's framing suggests. Meta has competitive models, unmatched consumer distribution and signs of commercial use. It is also investing before it has established that those assets can support returns proportionate to the cost.
A competitive model does not justify a $145 billion capital program by itself. Nor does a high spending level prove that Meta has little to show for it. Investors need evidence that model quality and distribution can become sustained usage, that usage can become revenue or protect the existing advertising business, and that those gains can exceed the long-term cost of infrastructure.
Meta may ultimately spend heavily on an AI strategy that fails to earn an acceptable return. The available evidence does not rule out that outcome. It does, however, make a declaration of spectacular failure premature. The more useful debate is not whether Meta has produced anything of value, but whether what it has produced can ever be valuable enough.
