Is Meta catching up with OpenAI?







Every funding round, every investor, and what the labs promised in return, in one document. Have a look ↓
Summary
Yes, Meta is catching up with OpenAI, and it has already reached roughly the same level in frontier AI model performance. OpenAI still has a substantial lead in dedicated AI users, enterprise customers, and revenue generated directly from AI products.
Meta's research comeback has been remarkably fast. After Llama 4 struggled, the company reorganized its AI operations and improved from an Artificial Analysis Intelligence Index score of 31 with Muse Spark to 56 with Muse Wonder Full in roughly six months, slightly ahead of GPT-6 Astra's 53. Going from 31 to 56 in six months is a big move, and we go through what has changed over the past year in our frontier AI labs report.
There is a catch to that benchmark victory. Meta's strongest model, Wonder Full, is restricted to Muse Code subscribers, while the broadly API-accessible Wonder scores 52. Meta can now match OpenAI's frontier capabilities, but its best performance is not available to every developer.
Coding performance reveals a split between different strengths. Wonder Full leads Astra on demanding long-running software-engineering benchmarks, including SWE-Marathon, while Astra performs better on independent terminal-based evaluations. Neither company has a universal technical advantage.
Muse's consumer launch has been unusually successful, with more than 10 million registrations in approximately six weeks and 2.5 million daily active users. But strong early retention and aggressive advertising are arriving together, making it difficult to separate genuine demand from the effects of Meta's launch campaign.
The user comparisons favor OpenAI once engagement is considered. ChatGPT's broader ecosystem has more than one billion weekly users, while Meta's earlier one-billion-user milestone measured monthly AI activity across existing applications. Those figures describe very different levels of habitual usage.
Meta's biggest long-term advantage may be distribution rather than model quality. Its applications reach 3.64 billion people daily, giving it opportunities to introduce AI through messaging, shopping, advertising, and smart glasses without asking everyone to adopt a separate chatbot.
OpenAI's enterprise lead is harder to overcome. It already serves more than 2.5 million businesses, while Meta's new Enterprise Platform is still establishing itself. Existing advertising relationships could help Meta attract smaller businesses, but they do not automatically translate into enterprise software adoption.
Meta's pricing advantage is less straightforward than advertised. Muse Wonder's API token prices are 60% lower than Astra's, yet its greater token consumption produces a roughly 39% higher average cost per evaluated task. Meta is much faster at generating output, creating a real trade-off between speed and total cost.
Meta can afford a prolonged fight. Its enormous advertising business supports planned annual capital expenditures of $135 billion to $145 billion, and its early AI shopping activity gives it another potential route to monetization. OpenAI, meanwhile, has already built the stronger standalone generative AI business.
Our assessment is that Meta's technological catch-up has largely happened, while its commercial catch-up is still beginning. The decisive questions are whether Muse users remain active after the marketing push, whether businesses adopt Meta's AI platform, and whether Meta can turn frontier-level models into a cheaper and more profitable service.
Written by the New Market Pitch teamWe track new markets so founders and investors can move faster. About us
70+pages
This is just a preview page. Don’t worry if a number looks old: the report gets refreshed, and you’ll always get the latest version.
Why is Meta suddenly a serious threat to OpenAI?
Meta is now a genuine threat to OpenAI because its latest AI models have reached the frontier, and its new Muse assistant is attracting users at an extraordinary pace.
The turnaround is striking. Meta's Llama 4 models, released in April 2025, struggled to compete with the strongest reasoning models from OpenAI, Anthropic, and Google. Mark Zuckerberg responded by reorganizing the company's AI research, investing approximately $14.3 billion in Scale AI, and recruiting its founder, Alexandr Wang, to help lead Meta Superintelligence Labs.
The results arrived much faster than many expected. Meta released Muse Spark in April 2026, followed by Muse Wonder and Muse Wonder Full in October. On the independent Artificial Analysis Intelligence Index, Meta progressed from a score of 31 for the original Muse Spark to 56 for Wonder Full. OpenAI's GPT-6 Astra scores 53 on the same index. One index score does not settle it; we separate who leads on benchmarks from who only claims it in our report on frontier AI labs.
Meta also has something most frontier AI developers lack: a massive existing consumer business. Its Facebook, Instagram, WhatsApp, and other applications reach 3.64 billion people daily, according to its latest quarterly results. Meanwhile, the standalone Muse assistant has already attracted more than 10 million account registrations in the United States and Canada.
Meta's technical progress has been much faster than its commercial catch-up. OpenAI has spent years building ChatGPT, developer APIs, coding tools, and enterprise relationships. Those advantages won't disappear because Meta released a better model.
But the competition has changed. Meta has the technology, the money, and an audience of billions. OpenAI can no longer count on Meta being a generation behind.
Has Meta actually caught up with OpenAI's best AI models?
Meta has essentially caught up with OpenAI in frontier AI performance, although GPT-6 Astra remains better at several important tasks.
Artificial Analysis gives Muse Wonder Full an Intelligence Index score of 56, compared with 53 for GPT-6 Astra at maximum reasoning effort. The API-accessible Muse Wonder scores 52, putting the two companies' commercially available models almost level.
Looking beyond the headline index reveals how close the competition has become. Muse Wonder and GPT-6 Astra both score 68% on Artificial Analysis's AutomationBench evaluation. They also tie at 81% on its long-context reasoning benchmark, while Meta leads on SciCode with 60% against OpenAI's 56%.
OpenAI performs better on several other tests. GPT-6 Astra scores 59% on Terminal-Bench 4.0 versus Wonder's 52%, and 55% on Humanity's Last Exam compared with 50%. Its factual reliability score on AA-Omniscience is also considerably stronger.
Another important distinction concerns availability. Wonder Full produces Meta's best results but is currently restricted to Muse Code subscribers. Developers using Meta's standard API receive Wonder, whose overall performance sits just below Astra.
Meta belongs in the same frontier-model category as OpenAI today. The better choice increasingly depends on the actual workload rather than which company leads a general intelligence ranking.
| Independent benchmark | Meta Muse Wonder | OpenAI GPT-6 Astra |
|---|---|---|
| Intelligence Index | 52 | 53 |
| AutomationBench | 68% | 68% |
| Terminal-Bench 4.0 | 52% | 59% |
| SciCode | 60% | 56% |
| Humanity's Last Exam | 50% | 55% |
| GDP.pdf | 31% | 31% |
| CritPt | 28% | 32% |
| Long-context reasoning | 81% | 81% |
| AA-Omniscience | 25 | 43 |
Figures use Artificial Analysis's maximum reasoning settings. Higher scores indicate better performance within each benchmark.
70+pages
Just a preview. If a number here looks old, no worries: your report comes with the latest data.
Did Meta's AI research overhaul really work?
Yes. Meta's AI research overhaul has already achieved its most immediate goal: producing models that can compete with OpenAI's best systems.
The changes began after Meta's earlier AI strategy failed to deliver the expected breakthrough. Llama 4 struggled to establish a clear advantage, while OpenAI and Anthropic continued advancing in reasoning and coding. Meta responded in June 2025 with its Scale AI investment and the recruitment of Alexandr Wang.
The company then reorganized its research operations around Meta Superintelligence Labs. Instead of concentrating primarily on releasing increasingly large Llama models, Meta put more resources into advanced reasoning, agents, and longer-running tasks.
The improvement is measurable. Muse Spark reached 31 on the Artificial Analysis Intelligence Index in April 2026. Approximately six months later, Muse Wonder Full scored 56, an increase of 25 index points.
That 25-point improvement is more useful than calculating a percentage increase, because the Intelligence Index is a composite benchmark rather than a linear measure of intelligence. Still, it's a huge jump for such a short period.
The restructuring also produced several distinct products. Muse Spark targets faster interactions, Muse Wonder handles more difficult tasks, and Muse Wonder Full pushes model performance further. Meta has begun separating efficient everyday models from its most advanced systems, much like its major competitors.
The research turnaround has worked. Whether Meta can repeat that progress across several model generations is another question, and one that won't be settled by a single successful launch.
Is Meta beating OpenAI at coding now?
Meta is beating OpenAI on some of the toughest published coding tests, particularly those involving long software-engineering tasks. OpenAI still leads on other developer benchmarks, so there is no universal winner.
Meta's Muse Wonder Full scored 79.0 on DeepSWE 1.1, compared with 74.1 for GPT-6 Astra, according to Meta's Wonder model card. The gap is larger on SWE-Marathon, where Wonder Full scored 57.5 against Astra's 42.5. Beating Astra on coding tests is one scoreboard; we sort out who is ahead and who is falling behind in our report covering frontier AI labs.
SWE-Marathon is especially interesting because it evaluates longer software-engineering work rather than isolated programming questions. Meta's 15-point advantage suggests that its strongest model can stay effective through complicated development tasks.
There are limitations to this comparison. Wonder Full was tested at a particularly high reasoning setting, and some competing models used different coding environments. The full performance gap therefore cannot be attributed to model quality alone.
Independent testing provides a useful counterexample. Artificial Analysis reports 59% for GPT-6 Astra on Terminal-Bench 4.0, compared with 52% for the standard Muse Wonder. That benchmark involves working through terminal-based technical tasks, where OpenAI still performs better.
Meta has also introduced Muse Code, putting its models into a developer workflow rather than offering them exclusively through a chatbot.
For developers, the choice has become genuinely competitive. Wonder Full looks particularly strong for sustained software engineering, while Astra retains advantages in some technical execution environments. Testing both on real repositories now makes more sense than choosing based on reputation.
Is Meta's Muse becoming a real competitor to ChatGPT?
Meta's Muse is already one of the fastest-growing AI applications ever launched in the United States, but ChatGPT remains far larger.
Mark Zuckerberg reported more than 10 million Muse account registrations in the United States and Canada within approximately six weeks of launch. Independent estimates from Sensor Tower put downloads at 10.9 million, including 10.2 million in the United States.
The speed is remarkable. Sensor Tower found that Muse reached 10 million US downloads in 40 days, a milestone achieved by only 11 other applications over the previous decade. Among nongaming apps, the other examples included Disney+ and Meta's Threads.
Muse also outperformed ChatGPT's original US mobile launch. Sensor Tower recorded 3.7 million US ChatGPT downloads during its first 40 days, compared with approximately 10 million for Muse. However, ChatGPT had roughly 15 million global downloads during its equivalent period, compared with Muse's 10.9 million, reflecting the different countries where the apps were available.
Early engagement looks promising too. Muse reached 2.5 million daily active users, up 15% from the previous week. Its seven-day retention rate was 30%, compared with a 19% average for ChatGPT, Claude, and Grok. After 14 days, Muse retained 23% of users against 16% for those competitors.
Those figures are impressive, although Meta is spending heavily to promote Muse. According to Sensor Tower, the application accounted for 39% of Meta's US advertising spending in October, up from 23% in September.
Meta has clearly figured out how to get people to try Muse. Keeping them is the harder part. The retention numbers are encouraging, but the application needs to maintain its momentum after the initial advertising push.
70+pages
This is a preview page. Spotted an outdated number? That’s normal here. The report you get is always the freshest version.
Does Meta already have more AI users than OpenAI?
OpenAI currently has the stronger documented position in dedicated AI usage, even though Meta reaches far more people through its existing applications.
Meta announced in May 2025 that Meta AI had reached one billion monthly active users, up from approximately 500 million in September 2024. The company had doubled its reported AI audience in about eight months by placing AI features throughout its social and messaging products.
OpenAI now reports more than one billion weekly active users across its products. Weekly usage is a more demanding measure than monthly usage, so the two billion-user figures cannot be treated as equivalent.
There is also a difference in how people interact with the products. Someone asking a question through WhatsApp may count toward Meta AI usage, while someone opening ChatGPT several times a week is showing a different level of engagement.
Muse gives Meta a clearer way to measure intentional AI adoption. Its reported 2.5 million daily active users represent a useful starting point, but remain small beside ChatGPT's worldwide audience.
We cannot calculate an exact market-share split because neither company provides directly comparable figures for every product, country, and usage frequency. The available data nevertheless puts OpenAI ahead in established AI usage. Meta has the larger audience to draw from, but that audience hasn't yet translated into comparable dedicated AI engagement.
Could Facebook, Instagram, WhatsApp, and Meta's smart glasses help it beat ChatGPT?
Absolutely. Meta has a distribution advantage that OpenAI will struggle to match: billions of people already use its applications, and the company also sells AI-enabled wearable devices.
According to Meta's latest earnings, 3.64 billion people use at least one application in its family each day. That gives Meta several ways to introduce AI into activities people already perform, including messaging, shopping, discovering content, and communicating with businesses.
Even modest adoption could produce enormous numbers. If 10% of Meta's daily audience used an AI assistant each day, that would represent approximately 364 million people. At 20%, it would be 728 million.
Those are hypothetical figures rather than forecasts. They illustrate why Meta can potentially reach a much larger audience without persuading everyone to install a separate application.
Its hardware strategy adds another opportunity. Ray-Ban Meta smart glasses already combine cameras, microphones, speakers, and conversational AI in a wearable product. Meta is also preparing deeper Muse integration with smart glasses and additional personal AI hardware.
These devices could make AI assistance more convenient for activities such as identifying objects, answering spoken questions, navigating unfamiliar places, and receiving contextual information.
OpenAI has a different advantage. People already open ChatGPT specifically to research, write, code, and complete difficult tasks. Meta still needs to turn casual interactions into that kind of repeated, intentional usage.
But there is a bigger possibility here. Meta could make AI so integrated into messaging, shopping, and wearable devices that many people use it constantly without ever thinking of opening a chatbot. OpenAI's lead in standalone assistants wouldn't necessarily protect it from that shift.
70+pages
Preview only. Some numbers may have changed since we made this page, but your report will be up to date.
Is Meta catching up with OpenAI in business AI?
Meta is now competing directly with OpenAI for business AI customers, but OpenAI remains much further ahead in actual commercial adoption.
OpenAI reports serving more than 2.5 million businesses worldwide. Enterprise customers accounted for over 40% of its revenue in April 2026, and its developer APIs were processing more than 15 billion tokens per minute.
Meta's response is the Enterprise Platform announced in September 2026. The platform brings together Muse, Meta Business Agent, Muse API, and Muse Code, allowing businesses to access Meta's models and develop AI-powered applications. Selling models to businesses through an enterprise platform is one business model among several; we compare which are winning in our frontier AI labs report.
Meta has an unusual advantage here. Zuckerberg said the company already helps hundreds of millions of businesses reach customers through its existing products. Its advertising relationships give it a direct route to small businesses that might benefit from automated customer support, sales assistance, and marketing tools.
However, selling advertising and selling enterprise AI software involve different requirements. Companies deploying agents across sensitive systems need reliable integrations, security controls, administrator tools, and predictable performance.
OpenAI already has an established customer base and products designed around those needs. Meta's Enterprise Platform is only beginning to build a comparable track record.
Meta could become especially competitive among smaller businesses already using Instagram, Facebook, and WhatsApp. For larger enterprise deployments, OpenAI has a much stronger position to defend.
Is Meta's AI cheaper than OpenAI's?
Meta's Muse Wonder is 60% cheaper than OpenAI's GPT-6 Astra on published input and output token prices, but its higher token consumption can make completed tasks more expensive.
Muse Wonder costs $4 per million input tokens and $20 per million output tokens. GPT-6 Astra costs $10 and $50 respectively, according to Artificial Analysis.
For developers processing large volumes of relatively straightforward requests, that pricing difference can be attractive. The comparison becomes less favorable to Meta when models use substantial amounts of reasoning.
Artificial Analysis measured approximately 54,000 output tokens per benchmark task for Muse Wonder, versus 27,000 for Astra. Muse Wonder also used around 29,000 reasoning tokens per task, compared with 17,000 for OpenAI's model.
As a result, the estimated average task cost was $4.52 for Muse Wonder and $3.26 for GPT-6 Astra. Meta's model was approximately 39% more expensive per completed benchmark task despite its much lower token rates.
Meta does have a speed advantage. Artificial Analysis measured around 176 output tokens per second for Muse Wonder, against 69 for Astra. Wonder's time to first token was also far shorter under the measured configurations.
The cheaper token price is therefore a slightly misleading selling point. Developers who care about total task cost may prefer Astra, while those prioritizing faster responses could find Muse Wonder more attractive. Real workloads will decide which advantage matters more.
| Pricing and efficiency measure | Meta Muse Wonder | OpenAI GPT-6 Astra |
|---|---|---|
| Input price per million tokens | $4.00 | $10.00 |
| Output price per million tokens | $20.00 | $50.00 |
| Cached input per million tokens | $0.20 | $1.00 |
| Weighted price per million tokens | $2.94 | $7.70 |
| Average output tokens per benchmark task | 54,000 | 27,000 |
| Average reasoning tokens per task | 29,000 | 17,000 |
| Average benchmark cost per task | $4.52 | $3.26 |
| Output speed | 176 tokens/s | 69 tokens/s |
| Context window | ~1.05 million tokens | ~1 million tokens |
Figures compare both models at maximum reasoning effort. Actual costs and speeds depend on workload and configuration.
70+pages
This is just a preview page. Don’t worry if a number looks old: the report gets refreshed, and you’ll always get the latest version.
Can Meta afford to spend more on AI than OpenAI?
Meta has one of the strongest financial positions in the AI industry, giving it the ability to fund enormous research and infrastructure projects without depending entirely on AI product sales.
The company's latest guidance puts 2026 capital expenditures at $135 billion to $145 billion, including principal payments on finance leases. Spread across the year, that represents approximately $370 million to $397 million per day.
Meta's second- and third-quarter capital expenditures reached $31.08 billion and $34.91 billion respectively, totaling nearly $66 billion in six months. This investment supports data centers, servers, networking equipment, and the computing capacity needed for advanced AI models.
Its advertising business provides the financial foundation. Meta generated $64.6 billion in third-quarter revenue, up 26% from $51.2 billion a year earlier.
Profitability has come under pressure, although the headline decline requires context. Meta's operating margin fell from 40% to 20% year over year, partly because of a $9.89 billion litigation charge. Without that charge, operating income would have been approximately $22.9 billion, representing an operating margin of about 35%.
OpenAI finances infrastructure through a different combination of customer revenue, external investment, and partnerships. Meta's capital spending therefore cannot be directly compared with every infrastructure commitment announced by OpenAI.
Meta's advantage is that its existing business can fund a very expensive AI strategy for years. Muse doesn't have to become profitable immediately for Meta to keep competing aggressively, and that's a luxury most AI developers don't have.
Has Meta abandoned open-source AI to compete with OpenAI?
Meta has become more selective about open-source AI, keeping its most advanced Muse Wonder models proprietary while continuing to release other powerful models with downloadable weights.
This is a major change from the original Llama strategy. Meta attracted developers by allowing them to download and customize model weights, giving companies greater control over how and where they deployed AI.
Muse Wonder and Wonder Full follow a more controlled approach. Wonder is available through Meta Model API and Muse Code, while Wonder Full is reserved for Muse Code subscribers. Neither is currently offered as an open-weight download.
At the same time, Meta released Muse Spark Open in October 2026. According to Meta AI Research, it has 816 billion total parameters, 42 billion active parameters, and a context window of approximately one million tokens.
Meta says the model can run using two Apple M5 Ultra Mac Studios and is also accessible through providers including Fireworks AI, Groq, Modal, and Together AI.
That combination reveals Meta's current strategy. Developers can still access powerful open-weight models, while Meta retains control over its very best systems and can charge for hosted access.
There is a trade-off. Some developers originally favored Meta because they could self-host its models rather than depend on OpenAI or another API provider. With Muse Wonder, Meta's frontier offering now comes with many of the same access restrictions.
Meta hasn't abandoned open weights, but it has clearly stopped treating openness as a requirement for every major model. When its strongest technology becomes a competitive advantage, the company wants to keep control of it.
Could Meta beat OpenAI in AI shopping?
Meta has a particularly strong opportunity in AI shopping because its existing business already connects consumers, advertisers, and merchants at enormous scale.
During Meta's latest earnings discussion, Zuckerberg said that more than one-quarter of Muse's weekly users were shopping with the agent. He also reported an average purchase value above $200 among users making purchases through Muse.
That is an encouraging early result for a newly launched assistant. It suggests people are willing to let Muse participate in purchases involving meaningful amounts of money, rather than using it only for simple questions. A quarter of Muse's weekly users shopping with the agent is an early opening; we look at where the biggest opportunities are in our frontier AI labs report.
Meta's advertising business gives the company an advantage in turning that activity into revenue. In the third quarter, advertising impressions rose 10% year over year, while average advertising prices increased 14%. Meta already has extensive experience connecting consumer attention with commercial outcomes.
OpenAI is moving into similar territory. Its ChatGPT advertising business reached a $1 billion annualized revenue run rate in less than 200 days, according to the company's August 2026 announcement.
The two companies have different starting points. OpenAI has built an assistant that people deliberately consult when making decisions. Meta has years of experience helping businesses reach consumers and is now adding an agent that can act on those consumers' behalf.
We still lack directly comparable transaction volumes, conversion rates, and revenue figures for AI-assisted purchases. But shopping could become one of Meta's strongest opportunities. It already has the merchants, the consumers, the advertising infrastructure, and plenty of data about commercial intent.
If Muse starts connecting those pieces effectively, Meta could build a shopping business that is very difficult for OpenAI to replicate from scratch.
70+pages
Just a preview. If a number here looks old, no worries: your report comes with the latest data.
Does OpenAI still have a bigger AI business than Meta?
OpenAI currently has a much larger standalone generative AI business, even though Meta is a substantially bigger and more profitable company overall.
OpenAI's position is supported by more than one billion weekly users, 2.5 million business customers, and a growing mix of consumer subscriptions, enterprise products, developer APIs, coding tools, and advertising.
Its enterprise business is becoming especially important. In April 2026, OpenAI said enterprise customers already generated more than 40% of revenue. That makes the company less dependent on individual ChatGPT subscriptions than it was during its earlier growth phase.
Meta's $64.6 billion in quarterly revenue puts it in a completely different corporate size category. However, the overwhelming majority of that revenue still comes from its established advertising and platform businesses.
Meta has not disclosed enough separate revenue figures for Muse subscriptions, Muse Code, the Model API, and its Enterprise Platform to establish the size of its new AI business.
There is an important complication. Meta already makes money from AI through advertising recommendations, content ranking, and automated marketing systems. The financial benefits of those systems are reflected within its existing business rather than reported as separate generative AI sales.
OpenAI leads in revenue generated directly from selling generative AI products and services. Meta has considerably more total revenue and several ways to benefit financially from AI without charging users directly.
It's a meaningful distinction. Meta can become a major winner from AI even if Muse never produces the kind of standalone subscription and API business that OpenAI has built.
What could stop Meta from overtaking OpenAI?
Meta's biggest obstacle is turning its new technical strength and early consumer popularity into a reliable business that people keep using. Three problems stand out in the current evidence: efficiency, trust, and enterprise adoption.
Efficiency is already measurable. Muse Wonder's lower API prices look attractive, but Artificial Analysis found that its higher reasoning-token consumption pushed the average evaluated task cost above GPT-6 Astra's. Meta will need to improve that relationship if it wants to compete heavily on price.
Trust is more difficult. Muse works across sensitive information, including messages, calendars, and personal documents. Security researchers have already identified privacy concerns involving publicly accessible shared Muse conversations, prompting Meta to suspend the sharing feature while investigating.
That issue is particularly serious for autonomous agents because they may send messages, make purchases, or interact with connected services on a user's behalf. A privacy mistake involving an agent can have consequences beyond an incorrect chatbot response.
Enterprise adoption presents another hurdle. OpenAI already has millions of business customers and established products designed for developers and organizations. Meta must convince businesses to adopt a new platform while continuing to support its existing advertising customers.
The company also needs to maintain its research progress. Muse Wonder Full's performance is impressive, but Meta has only recently reached the frontier again. OpenAI, Anthropic, and Google are continuing to release stronger models, and Meta will need to keep pace across multiple generations.
None of these problems looks insurmountable given Meta's resources. But the easiest part of the comeback may already be over. Building a competitive model is one thing; getting people and businesses to trust it, pay for it, and keep using it is harder.
70+pages
This is a preview page. Spotted an outdated number? That’s normal here. The report you get is always the freshest version.
Is Meta catching up with OpenAI, or has it already caught up?
Yes, Meta is catching up with OpenAI, and in frontier AI model performance it has largely caught up already. OpenAI still leads in established AI usage, enterprise adoption, and direct monetization.
Meta's technical turnaround is the strongest part of the story. After struggling with Llama 4, the company developed Muse Wonder Full, which now scores 56 on the Artificial Analysis Intelligence Index against GPT-6 Astra's 53. It also performs exceptionally well in several demanding coding and professional-agent evaluations.
The consumer results are promising. Muse attracted more than 10 million account registrations in approximately six weeks, while Sensor Tower measured 2.5 million daily active users and stronger early retention than several competing AI applications.
Meta also has financial and distribution advantages that few competitors can match. Its applications reach billions of people daily, its advertising business generates tens of billions of dollars each quarter, and its planned annual capital spending exceeds $135 billion.
Yet OpenAI remains clearly ahead as a standalone AI business. Its more than one billion weekly users, 2.5 million business customers, and established developer ecosystem represent years of accumulated adoption that Meta has not yet replicated.
Our assessment is that Meta has already achieved technological parity with OpenAI in several important areas. Its commercial catch-up is still at an earlier stage, and the size of that gap shouldn't be underestimated.
The next decisive evidence will come from Muse's retention after the initial marketing push, adoption of Meta Enterprise Platform, and whether Muse Wonder can become more cost-efficient on real tasks.
Meta has proved that it can build AI models good enough to challenge OpenAI. Turning them into an AI business of comparable scale is the much bigger test.
70+pages
Preview only. Some numbers may have changed since we made this page, but your report will be up to date.
Our methodology
This analysis examines whether Meta is catching up with OpenAI across frontier model performance, coding capabilities, consumer adoption, enterprise AI, model economics, distribution, and financial resources. We assessed these areas separately because having a competitive AI model does not necessarily mean having a competitive AI business.
For technical performance, we prioritized independent evaluations from Artificial Analysis, including its Intelligence Index, coding benchmarks, reasoning tests, output speed, and API cost measurements. We compared Muse Wonder with GPT-6 Astra under maximum reasoning settings and treated Muse Wonder Full separately where its availability or testing conditions differed.
We also reviewed Meta's published model documentation and OpenAI's model announcements to understand the differences between headline benchmark performance and the capabilities developers can actually access. Meta's reported software-engineering results were considered alongside independent terminal-based evaluations rather than treated as directly interchangeable measurements.
To assess consumer adoption, we used Meta's disclosures about Muse registrations and Meta AI usage, together with Sensor Tower estimates reported by EMARKETER. We distinguished account registrations, downloads, daily active users, weekly users, monthly users, and retention rates. These measures capture different behaviors and cannot be combined into a single market-share figure.
For commercial adoption, we compared OpenAI's disclosed business customer base, enterprise revenue contribution, and API activity with Meta's Enterprise Platform announcement and its existing relationships with advertisers and merchants. We treated Meta's distribution reach as an opportunity to acquire AI users, rather than evidence that those users have already adopted Muse.
We assessed pricing using published input and output token rates, measured token consumption, average evaluated task costs, and generation speeds. As discussed above, per-token pricing and cost per completed task can produce different rankings, particularly when models use substantially different amounts of reasoning.
Our financial assessment draws on Meta's quarterly financial releases, capital expenditure guidance, and operating results, alongside OpenAI's disclosures about its user base, enterprise business, and advertising. We separated Meta's total corporate revenue from revenue attributable specifically to generative AI products, since its existing advertising operations account for most of its reported business.
We also examined Meta's open-weight strategy, smart glasses, shopping activity, and reported privacy issues to understand where the company could gain an advantage and where its expansion might face resistance. Potential distribution and monetization benefits were kept separate from outcomes that have already been measured.
Key sources used for this analysis include: Scale AI's announcement of Meta's strategic investment, Meta's Muse Spark launch, Meta AI Research on Muse Wonder, Artificial Analysis's Muse Wonder evaluation, Artificial Analysis's Muse Wonder versus GPT-6 Astra comparison, OpenAI's GPT-6 Astra documentation, Meta's Muse assistant announcement, EMARKETER's reporting on Sensor Tower adoption data, Meta's third-quarter financial results, and Meta's second-quarter financial results.
Additional sources include: OpenAI's disclosure of its user and business customer base, OpenAI's enterprise AI update, Meta's Enterprise Platform announcement, Meta AI Research on Muse Spark Open, the Muse Spark Open model card on Hugging Face, Meta's AI glasses announcement, OpenAI's ChatGPT advertising update, and WIRED's reporting on Muse conversation-sharing privacy concerns.
Our conclusions give greater weight to independently measured model performance, observed user engagement, disclosed business adoption, and operating economics than to hypothetical adoption scenarios. We assess Meta as having reached frontier-level technical competitiveness while OpenAI retains its lead in established, directly monetized AI usage.
Written by the New Market Pitch teamWe track new markets so founders and investors can move faster. About us









