Key Takeaways

• Meta Superintelligence Labs has closed part of the gap on rivals but still trails OpenAI, Anthropic and Google on frontier model quality and on model revenue, which it only began collecting in July 2026.
• Its clearest advantages are distribution and price: Muse Spark 1.1 launched at $1.25 input and $4.25 output per million tokens, roughly a quarter of what Anthropic and OpenAI charge for comparable models.
• MSL was founded on 30 June 2025 and organised into four teams — TBD Lab, FAIR, Products and Applied Research, and MSL Infra — under chief AI officer Alexandr Wang.
• Meta paid over $14 billion for a stake in Scale AI and to bring Wang across, then cut roughly 600 roles from the division in October 2025 while sparing TBD Lab.
• Muse Spark 1.1 tops professional and scaled tool-use benchmarks — 54.7 on JobBench, 88.1 on MCP Atlas — but trails on pure coding: 61.5 on SWE-Bench Pro against Claude Opus 4.8’s 69.2.
• OpenAI reached roughly $25 billion annualised revenue by mid-2026; Anthropic reported $30 billion by April 2026, with a more conservative read near $22 billion.
• Meta’s 2026 capital expenditure guidance rose to $125–145 billion, against Alphabet’s $175–190 billion and Amazon’s projected $200 billion.
• The public preview of Meta’s API is US-only with a waitlist and absent from third-party routers, limiting adoption relative to globally available rivals.

Meta Superintelligence Labs is back in the race but not leading it. A year after Mark Zuckerberg spent over $14 billion to install Alexandr Wang and rebuild Meta’s AI organisation, the division has shipped three proprietary models and opened a paid API — real progress from a starting point where Llama 4 was widely judged a disappointment. On frontier capability, though, it sits behind OpenAI, Anthropic and Google, and it has no model revenue history to speak of.

Where MSL is genuinely differentiated is on the two things Meta already owned before any of this began: billions of users to distribute into, and an advertising business large enough to fund frontier compute without selling API access. That combination lets Meta price its models at roughly a quarter of Anthropic’s and OpenAI’s rates, which is a strategic weapon the pure-play labs cannot easily answer.

What Meta Superintelligence Labs Actually Is

MSL was established on 30 June 2025 in Menlo Park as an umbrella for Meta’s foundation models, research, and AI products, employing around 3,000 people. In August 2025 it reorganised into four teams: TBD Lab, the secretive frontier group Wang leads directly; FAIR, the long-standing fundamental research unit; Products and Applied Research, run by former GitHub chief executive Nat Friedman; and MSL Infra. Shengjia Zhao, a ChatGPT co-creator hired from OpenAI, became chief scientist.

The build-out was expensive and disruptive in equal measure. Meta reportedly offered individual researchers compensation packages in the hundreds of millions. In October 2025 the division cut roughly 600 positions from FAIR, product AI, and infrastructure while continuing to hire into TBD Lab, on the reasoning that the older organisation had become too bureaucratic to move quickly. Yann LeCun, Meta’s chief AI scientist, left in November 2025 after being placed under Wang. Several senior hires departed within months of joining, a number of them returning to OpenAI.

The Model Lineup

MSL’s output arrived in a compressed three-month window in 2026.

Muse Spark (8 April 2026) was the first model from the new organisation and Meta’s first proprietary foundation model, breaking with the open-weight Llama tradition. Built natively multimodal after a nine-month rebuild of Meta’s pretraining stack, Meta reported reaching Llama 4 Maverick’s capability level using over an order of magnitude less training compute. It shipped as a select-partners preview.

Muse Image (7 July 2026) was the lab’s first media generation model, replacing image technology Meta had licensed from Midjourney and Black Forest Labs. It generates agentically — searching the web and executing code before producing output — and ranked second on the Arena image leaderboard at launch. A Muse Video preview shipped alongside it.

Muse Spark 1.1 (9 July 2026) turned the research model into a developer product and opened the Meta Model API, the first time Meta has charged for access to its own model. Wang described the pricing as aggressive and attractive relative to comparable offerings from Anthropic and OpenAI.

Watermelon is in training. Wang told an internal town hall that the model, the successor to Muse Spark’s codename Avocado, has caught up with OpenAI’s GPT-5.5 on closely followed benchmarks, though he did not specify which.

Price: Meta’s Sharpest Weapon

Model	Input per 1M	Output per 1M	Context
Meta Muse Spark 1.1	$1.25	$4.25	1M (some listings cite 262K)
xAI Grok 4.5	$2.00	$6.00	500K
Claude Sonnet 5	$2.00 intro / $3.00 standard	$10.00 intro / $15.00 standard	1M
Claude Opus 5	$5.00	$25.00	1M
OpenAI GPT-5.5 (reported)	~$5.00	~$30.00	—
Claude Fable 5	$10.00	$50.00	1M

On output tokens — where agentic workloads spend most of their budget — Meta undercuts Opus 5 by roughly six times and GPT-5.5 by seven. Cached input at $0.15 per million sits below most competitors’ standard uncached input rate. New accounts get $20 in credits. For a workload generating a million output tokens daily, that is $4.25 against $25 on Opus, or roughly $127 a month against $750.

Note the caveat: Muse Spark 1.1 is a reasoning model with an effort parameter, and thinking tokens bill at the output rate. As with every reasoning model, the sticker price and the invoice can diverge. Our comparison of subscription pricing across the major assistants covers the consumer-facing half of the same competition.

Capability: Where MSL Leads and Where It Does Not

The benchmark picture is unusually clean. Muse Spark 1.1 leads on tool use and orchestration — 54.7 on JobBench and 88.1 on MCP Atlas, both professional and scaled tool-use measures — which is exactly the capability that matters for agents that drive real software. It concedes raw coding accuracy: 61.5 on SWE-Bench Pro against Opus 4.8’s 69.2, and 53.3 on DeepSWE against GPT-5.5’s 67.0. Long-horizon agentic work remains weaker than either rival.

That profile suits Meta’s product strategy more than it suits developer adoption. An assistant that plans across apps and services, books things, and drives commerce needs tool-calling reliability more than it needs to resolve GitHub issues. Whether developers value the same thing is a different question, and the benchmark that answers it has not been run yet.

One historical caution applies. Meta previously published Llama 4 benchmark results using specialised unreleased variants tuned for specific tasks, then acknowledged the general release did not match. Independent verification of Muse Spark’s numbers carries more weight than usual here.

The Financial Comparison

	Meta / MSL	OpenAI	Anthropic	Alphabet
Model revenue	API opened July 2026; no history	~$25B annualised, mid-2026	$30B reported April 2026 (~$22B conservative)	Not separately disclosed
Valuation / market cap	Public company; AI not separately valued	$852B post-money	$965B after May 2026 Series H	~$4.2T
2026 capex guidance	$125–145B	Stargate: $500B multi-year commitment	$200B / 5 years to Google Cloud for 5 GW	$175–190B
Funding model	Advertising cash flow	External capital; ~$14B projected 2026 loss	~$72B raised; projected $14B 2026 loss	Operating profit plus $84.75B equity raise
Distribution	Facebook, Instagram, WhatsApp, Ray-Ban glasses	ChatGPT, ~20M paid seats	Enterprise, API, Claude Code	Search, YouTube, Android, Cloud

The structural difference is visible in the funding row. OpenAI and Anthropic are both burning capital heavily against fast revenue growth, and the four largest US cloud providers hold roughly $2.1 trillion in revenue backlog with about half of it committed by those two companies. Meta funds its models from advertising, which removes the pressure to monetise inference at any particular margin — and, by the same token, removes the discipline that comes with having to.

Alphabet occupies a category of its own, owning models, research, cloud, silicon, and distribution simultaneously. That vertical position is the reason its market capitalisation ran up roughly 91% over the trailing year. Meta owns distribution and is buying compute; it does not own the silicon layer, which is where our piece on OpenAI’s push to control its own infrastructure traces the same logic playing out at a rival.

The Access Problem

Price advantages only convert if developers can use the product. Meta’s public preview is US-only, gated behind a waitlist, and deliberately kept off third-party routers such as OpenRouter. Developers in Europe and elsewhere cannot access it at all. Meanwhile the Claude and GPT models it benchmarks against are already available globally, and Anthropic and OpenAI both offer batch discounts and mature caching that narrow the effective price gap for anyone who uses them properly.

There is also the matter of what Meta gave up. The Llama family reached 1.2 billion downloads and roughly a million a day by early 2026, offering self-hosters an estimated 88% cost reduction against proprietary APIs. Closing the frontier line trades that goodwill for revenue, and the developer reaction has been sceptical. Meta has said future versions may be open-sourced without committing to a date.

The Honest Verdict

Meta Superintelligence Labs is a credible fourth entrant rather than a leader. It ships fast, prices aggressively, and reaches more people than any competitor through apps that billions already open daily. It does not hold the frontier on any capability that matters to serious builders, its access is geographically restricted, and it has spent a year of severe internal turbulence — a 28-year-old chief AI officer, repeated reorganisations, senior departures — to reach parity-adjacent rather than parity.

The comparison to make in six months is not benchmark scores but paying users. Zuckerberg now has a model; the open question is whether Meta can sell AI as a product rather than only using it to strengthen advertising. Watermelon’s release, and whether it ships with global API availability, will answer more than any leaderboard. For the origins of the whole effort, our account of Meta’s $15 billion Scale AI investment covers the deal that started it.

Figures are anchored to 28 July 2026. Revenue and valuation numbers for private companies are reported estimates and vary between sources. This article is informational and not investment advice.

If you are interested in this topic, we suggest you check our articles:

• Meta’s $15B Bet on Scale AI: Why Data is the New Gold?
• xAI vs OpenAI vs Anthropic: Which AI Lab Wins in 2026?
• OpenAI’s Infrastructure Push: Becoming Its Own Cloud Provider
• Claude Opus 4.6 vs 4.7 vs 4.8: Which Model Wins?
• 2026 AI Subscription Prices: Gemini vs ChatGPT vs Claude

Sources: CNBC, VentureBeat, The Agent Report, DataCamp, Axios, Meta AI blog

Written by Alius Noreika