Meta spent the better part of a decade insisting that open weights were the future of artificial intelligence. On July 9, 2026, it quietly closed that chapter. Meta Superintelligence Labs shipped Muse Spark 1.1 — a multimodal reasoning model tuned for agentic work, wrapped around a 1-million-token context window — and paired it with the Meta Model API, the company's first-ever paid developer endpoint. For the first time, outside developers can rent a frontier Meta model by the token instead of downloading the weights for free.
The move drops Meta squarely into the commercial arena occupied by Anthropic, OpenAI and Google, and it does so on price. The API lists at $1.25 per million input tokens and $4.25 per million output tokens, with $20 in starter credits. Mark Zuckerberg framed that as roughly a quarter of what rivals charge for comparable frontier models. It is, in other words, a land grab dressed as a model release.
What actually shipped
Muse Spark 1.1 is the second model out of the Superintelligence Labs group now run by 28-year-old Chief AI Officer Alexandr Wang, following the original Muse Spark in April. The headline capability is agentic execution rather than raw chat quality. Zuckerberg described the model bluntly in a rare public post: "Muse Spark is strongest at agentic performance, tool use, and computer use. It does well on long-running tasks with 1M token context window, can delegate execution to sub-agents running in parallel, and is trained to use computer interfaces on desktop, mobile, or browser."
That description maps to three distinct bets. The first is the 1M-token context, which the model manages through active compaction — retrieving earlier work, dropping noise and preserving the steps it will need later, a direct answer to the familiar failure mode of agents that forget what they were doing mid-refactor. The second is parallel subagent delegation: Muse Spark 1.1 can act as a lead agent that farms tasks out to constrained subagents running simultaneously, or serve as one of those subagents and escalate when it hits a wall. The third is native cross-surface computer use, with the model trained to operate desktop, browser and mobile interfaces and to choose between writing scripts, taking direct interface actions, or batching actions depending on the job.
The benchmark story is genuinely mixed, and worth reading carefully. Wang touted state-of-the-art results on agent and tool-use evals including Harvey's Legal Agent Bench, TaxEval, MedScribe, MCP Atlas, Finance Agent v2 and HealthBench Professional, along with out-of-distribution tests where Meta says it edges Anthropic's Claude Opus 4.8 and xAI's Grok 4.5. But on the coding benchmarks developers actually argue about, Muse Spark 1.1 trails: it scores 61.5 on SWE-Bench Pro, behind Opus 4.8 at 69.2 though ahead of GPT-5.5 at 58.6, while GPT-5.5 leads on Terminal-Bench 2.1 and DeepSWE 1.1. The Terminal-Bench number drew immediate scrutiny — Meta ran the eval with 6-core, 8GB-RAM compute caps that critics noted don't match the benchmark's standard configurations, making the headline figure the largest asterisk in the release.
Why this matters
The strategic shift is bigger than any single benchmark. Meta built its AI reputation on Llama's open weights; charging for API access ends the free-weights era for its flagship line and signals that the company now sees agents — not downloadable models — as the product worth monetizing. The aggressive pricing is the tell. Undercutting Anthropic and OpenAI by roughly 75% is a play for developer volume and mindshare in the agentic-coding and computer-use market that Anthropic, in particular, has come to dominate.
There are real caveats. Access is gated: US developers can start immediately, with a waitlist for everyone else, and Meta briefly listed the model on OpenRouter before restricting broader availability. The community consensus is measured. As one developer guide put it, Muse Spark 1.1 is "not a switch now for most teams" but rather a "watch closely, prototype when you get in" release — a credible frontier contender whose main near-term effect may be pricing pressure on the rest of the field.
What to watch next
Three things. First, whether Meta lifts the geographic waitlist and posts reproducible benchmark configs to quiet the Terminal-Bench dispute — transparency here will shape how seriously enterprise buyers take the SOTA claims. Second, how fast the computer-use and parallel-subagent features hold up outside cherry-picked evals, since long-horizon reliability is where agentic models tend to break. And third, whether rivals respond on price. If a company with Meta's balance sheet is willing to sell frontier agents at a quarter of the going rate, the more interesting question is not how good Muse Spark 1.1 is today, but how long anyone else can charge four times as much.
"Muse Spark is strongest at agentic performance, tool use, and computer use. It does well on long-running tasks with 1M token context window, can delegate execution to sub-agents running in parallel, and is trained to use computer interfaces on desktop, mobile, or browser."- Mark Zuckerberg, CEO, Meta