For three years, Meta had a clear persona in AI: the king of open source. The Llama series swept the open community, and Zuckerberg repeated in interviews that “AI should be open like Linux.” Developers used the free models and figured Meta was the most conscientious of the Silicon Valley giants.
On July 9, Meta released Muse Spark 1.1. It looks like another model update, but several details differ: pricing at $1.25 per million input tokens and $4.25 per million output — slightly above GPT-5.6 Luna but below Claude Sonnet 4.6. Zuckerberg posted on X for the first time in three years to promote it, calling Spark “an extremely low-priced but capable agentic and coding model.”
That alone is no secret. What deserves attention is that Meta did three things at the same moment: First, Muse Spark will replace parts of the Llama models inside WhatsApp, Instagram, Facebook chatbots, and the Meta smart-glasses line. Second, Meta is training a model codenamed Watermelon that already matches GPT-5.5 on key benchmarks. Third, an open-source variant of Muse Spark is in development — with the team refusing to say when it ships.
Together, these three tell a story that is not “another model update.” The story is: Meta’s AI strategy is pivoting from “open-source everything” to “API first.”
Open Source Was a Strategy, Not a Faith
Understanding the pivot starts with why Meta open-sourced at all. When Llama opened in 2023, Meta’s logic was simple: in the AI race, Meta was the chaser. OpenAI had GPT-4’s first-mover advantage, Google had DeepMind’s accumulation — what did Meta have? User data and application scenarios. Open source was the fastest path to community contribution, developer ecosystem, and accelerated iteration.
Not charity — investment. Two years later, that investment needs to start paying out. Meta’s AI infrastructure spend is in the hundreds of billions — 2025 capital expenditure alone exceeded $37 billion. The Llama series won developer goodwill but brought Meta no direct revenue. Value capture from open models is a long game, and Meta’s shareholders cannot wait that long. Muse Spark 1.1’s launch is, in essence, the bill for value capture arriving. Zuckerberg posting on X was not a whim: his last post was July 2023 when Twitter renamed itself X. Three years silent, and the first post sells product — that signal outweighs any strategy memo.
From the Capability Race to the Commercialization Race
Muse Spark 1.1’s pricing deserves scrutiny. $1.25/$4.25 per million input/output tokens — above GPT-5.6 Luna ($1.10/$3.60), well below Claude Sonnet 4.6 ($3.00/$15.00). Meta positioned its model as “cheaper than Google, more affordable than Anthropic, slightly pricier than OpenAI but not worse.” Not a random price point — it reflects a structural change: AI competition has entered its second stage.
Stage one (2022-2025) was the capability race: whoever’s foundation model was strongest won, and every launch was reported as “how much better than the last.” Stage two (2025-) is the commercialization race: the capability gaps are narrowing — GPT-5.6, Claude Sonnet 4.6, Grok 4.5, Muse Spark 1.1 — and users can no longer tell by intuition who is stronger. Competition shifts from “who is smarter” to “who is cheaper, easier to use, better integrated.” It parallels the smartphone market’s watershed around 2010: after the iPhone 4, hardware capability stopped being the core moat, and app ecosystems, pricing strategy, and brand positioning decided the winners. Meta entering at this stage is well timed — it need not crush anyone on capability, only be “good enough” on price and service, then use its app ecosystem to put the model in front of billions of users.
Closed Models Earn, Open Models Claim Territory: the Dual-Track Weapon
The open-source Muse Spark variant is in development, with Meta refusing to commit to a date. The non-commitment is more informative than a promise would be. It implies a dual-track strategy: the closed API earns money — Muse Spark 1.1 charges through the API for enterprise developers, priced aggressively but sensibly so that SMBs and indie developers buy without agonizing. The open models claim territory — Llama keeps going open-source but will gradually be repositioned as the “entry-level” option; the community gets models that are “sufficient but not cutting-edge,” while the real commercial capability stays wrapped in the API.
The strategy is not new — Red Hat monetized open Linux through paid support; MongoDB opened its database and monetized the enterprise edition. But Meta’s scale is another order of magnitude: it simultaneously owns the open ecosystem, billions of users, and hundred-billion-level infrastructure. OpenAI lacks a user ecosystem (ChatGPT’s 400 million MAU is application layer, not a social graph). Anthropic lacks infrastructure scale. Google has an ecosystem but a wavering open-source stance. Only Meta can play both the “open source” and “API” cards and let developers choose.
What the Frame Explains
“AI competition has entered stage two” explains several live events. Anthropic’s Fable 5 free period keeps extending — not charity but standard commercialization-race moves: use free access to build user habits, then monetize the API. Fable 5 may be strong, but Anthropic’s pricing lacks competitiveness, so extending free access is how it defends the user base. OpenAI removed Codex usage limits — same logic. GPT-5.6 Luna’s $1.10/$3.60 already makes it one of the cheapest premium models; removing limits gets developers addicted. Once your dev flow binds to Codex, switching cost becomes astronomical. Grok 4.5’s reputation turned upward — Musk’s “Grok 4.5 beats Opus 4.8” is marketing, but the logic behind it is real: when capability gaps shrink below the perception threshold, experience feel and brand preference become decisive.
Three phenomena, one frame: model capability is no longer the single variable. Pricing, experience, ecosystem, and brand — the “commercial layer” — are becoming AI competition’s main battlefield.
What To Do
Developers: look past benchmarks when choosing models. Run the numbers for your scenario — API cost, latency, context length, toolchain integration. Meta’s $1.25/$4.25 may beat GPT-5.6 Luna for coding workloads, since Spark is optimized for agentic coding. AI founders: note the trend — foundation models are becoming infrastructure commodities, so your differentiation cannot rest on “which model you use,” only on “what you build with it.” Meta’s Watermelon is already chasing GPT-5.5, and that gap-closing will accelerate. Enterprise CTOs: start tracking model switching cost. Do not bind your application architecture to one vendor’s API. Meta’s entry means the price war will continue, and teams with flexible architectures will capture every wave of price cuts.
References: TechCrunch: Meta enters the crowded AI coding battle with Muse Spark 1.1 · InfoQ: Zuckerberg’s first post in three years promotes it — Meta’s strongest agent model enters coding
