Meta shipped its answer to a question that has hung over the company since it spent $14.3 billion for a slice of Scale AI: can Alexandr Wang actually build a frontier model? On Thursday, July 9, Meta Superintelligence Labs released Muse Spark 1.1, an upgrade to April's debut model — and, for the first time, opened it to developers through a paid API. The pitch is not that Meta has caught the leaders. It's that Meta will undercut them on price.
What actually shipped
Muse Spark 1.1 is a refresh of the reasoning model Meta launched on April 8, focused on the two things enterprises are actually paying for in 2026: coding and agentic tasks. Wang, Meta's first-ever chief AI officer, told Axios those were the explicit priorities for this release. The model also handles multimodal work like video captioning and general reasoning, and it now powers the "thinking" mode inside the Meta AI app and website.
The bigger structural change is the Meta Model API, now in public preview. Until this week, Muse Spark was a consumer feature buried inside Facebook, Instagram, WhatsApp, and Meta AI. Developers had no way to build on it. That's now changed, and the pricing is the headline.
Muse Spark 1.1 costs $1.25 per million input tokens and $4.25 per million output tokens.
Those numbers matter because of what they undercut. As Axios noted, that's cheaper than xAI's newly released Grok 4.5 and less than Anthropic's Opus tier. Zuckerberg framed it bluntly on X: "Our focus is on delivering strong agentic and multimodal models at very low cost. More to come soon." This is a deliberate wedge — compete on cost-per-token while the frontier labs fight over the top of the benchmark charts.
The benchmark claims deserve a skeptical read
Meta says Muse Spark 1.1 beats Google's latest Gemini on coding and reasoning benchmarks, and surpasses older OpenAI and Anthropic models on some verticals. Read that sentence twice. The word doing the heavy lifting is older.
Meta pointedly did not compare Muse Spark 1.1 against the newest flagships — Anthropic's Mythos 5 and Fable 5, or OpenAI's GPT-5.6. On at least one coding metric, the open-source Terminal-Bench leaderboard still shows Muse Spark trailing those top models. Benchmarking against a competitor's previous generation is a familiar move, and it tells you where a model really sits: strong, but not yet at the frontier.
There's history here, too. In April 2025, Meta faced accusations that it gamed benchmark results for a Llama 4 release. A former Meta AI executive denied training on test sets at the time. The company carries that credibility question into every performance claim it now makes, and it should — extraordinary benchmark claims from any lab deserve independent verification, not a press release.
Why Meta is doing this
Meta's AI story over the past year has been one of expensive reinvention. The company originally committed to open-source models with Llama. Then it pivoted, spending $14.3 billion in 2025 for a 49% non-voting stake in Scale AI and installing Wang — Scale's founder — to run a reorganized unit now branded Meta Superintelligence Labs. Muse Spark, unlike Llama, is proprietary and closed.
The reorganization was not smooth. Reporting throughout late 2025 described staff whiplash as Wang and Zuckerberg rebuilt the AI org from the ground up, and a new Applied AI unit pulled engineers into data-collection work some viewed as drudgery. Muse Spark 1.1 is the first real evidence that the reshuffle is producing shippable models at a steady cadence rather than just headcount churn.
That cadence is worth noting. In a single week, Meta released Muse Image and Muse Video — its first image and video generation models from the new lab — followed two days later by Muse Spark 1.1. The image launch was not without controversy: Meta let users apply AI editing effects to photos other people had posted publicly on Instagram, with no explicit permission required, which drew immediate backlash. Shipping fast and cleaning up later is a pattern worth watching.
What's next: a model called Watermelon
The most important detail may be the one Meta is quietest about. Muse Spark 1.1 is explicitly not the big leap the company is chasing. Wang and Zuckerberg have signaled a far more powerful model — code-named Watermelon — still in training, using vastly more compute, due later this year.
That reframes Muse Spark 1.1. It's a placeholder that keeps developers engaged and generates API revenue while the compute-heavy flagship bakes. Wang's longer-term vision, as he described it to Axios, is a genuinely agentic Meta AI that could plan a vacation or a party and take actions on its own — leaning on the one asset no rival can copy: billions of users and the data Meta holds on them.
Should you build on it?
For developers, the calculus is narrow but real. If you're running high-volume, cost-sensitive workloads where a previous-generation-class model is good enough — bulk classification, summarization, agentic loops that tolerate retries — Muse Spark 1.1's pricing is genuinely attractive. For frontier coding or reasoning where quality is the constraint, the leaderboards say you're still better served elsewhere for now.
The strategic read is clearer than the technical one. Meta is not trying to win the benchmark war this quarter. It's trying to make itself unignorable on price while Watermelon trains, and to convert its consumer reach into a developer platform. Whether that works depends on Watermelon delivering — and on Meta rebuilding the trust its benchmark claims keep spending.
The Bottom Line
Muse Spark 1.1 is a competent, aggressively priced model that Meta is smart to open to developers — but the benchmark framing is a tell. Beating a rival's older models at a lower price is a viable business strategy, not a frontier breakthrough. The real test is Watermelon, and it hasn't shipped. For now, treat Muse Spark 1.1 as a cost play worth trying, and treat every performance claim as something to verify yourself.
