Meta shipped Muse Spark 1.1 in the middle of July. The benchmarks are good, the price is low, and the reviews are hands-on and cheerful. Almost every write-up buries the one fact that actually matters: there is no weights file. Muse Spark 1.1 is Meta's first paid, API-only model. You pay per token to call it, the way you pay OpenAI and Anthropic, and you cannot download it and run it yourself.
For any other company that would be a footnote. For Meta it is the story.
The company that made "open" a weapon#
Llama was not just a model line. It was a strategy and an argument: that the way to win in AI was to give the weights away, seed an ecosystem, and let a thousand fine-tunes bloom while your competitors metered access by the token. "Download the weights" became a rallying cry, a moral position, and a moat all at once. Meta spent three years being the loudest voice in the room for open models.
Muse Spark 1.1 quietly sets that voice down. It is pay-as-you-go — roughly $1.25 per million input tokens and $4.25 per million output, with $20 of free credits to get you hooked — on a one-million-token context window with active compaction. It is an API. It is closed. And it is, by Meta's own framing, the best model the company has ever built for the thing that matters most commercially right now: doing work.
The open-model era isn't ending. But the frontier of agent capability is going closed even at its most open vendor.
The benchmark is the business plan#
Watch which number Meta chose to lead with. Not a general-reasoning score, not a coding crown — Muse Spark 1.1 is middling on pure coding, and on OSWorld-Verified computer use it posts 80.8, behind Claude Opus 4.8's 83.4. The headline is JobBench, which measures professional tool use: multi-step, tool-heavy tasks of the kind a business wants automated. There, Muse Spark 1.1 scores 54.7 against Opus 4.8's 48.4 and GPT-5.5's 38.3. That is not a nose ahead. That is a lap.
A benchmark a company leads with is a customer it is naming. Meta did not build the smartest model; it built the one that finishes office work, and it priced that model to be run at volume. Researchers download weights. Businesses buy throughput. Muse Spark 1.1 is aimed squarely at the second group, and its whole shape — cheap per token, huge context, tuned for tool use, no weights to bother self-hosting — is designed to make automating a workflow a line item instead of a project.
Why even Meta closed the door#
The uncomfortable read is the simplest one. Base models have commoditized; a capable general model is nearly free and getting freer. Agent capability has not. The ability to reliably chain tools, hold a million tokens of task context, and drive a browser to completion is where the willingness-to-pay still lives — and it monetizes per token, per task, per seat. Giving that away as weights would be giving away the one thing customers will still pay for. So Meta didn't.
If you want the pricing tactics — when Muse Spark's quarter-price tier actually lowers your bill and when it doesn't — we covered that in Muse Spark's API is a quarter of the price, and where it fits in a routing table in Terra vs Muse Spark vs Grok.
What it means if you're building#
Two things, and they pull against each other.
The opportunity: Muse Spark 1.1 is a genuinely strong, genuinely cheap option for tool-heavy professional automation. If that's your workload, it belongs in your routing table today — benchmark it against your own task distribution first, because leaderboard wins rarely transfer cleanly, but the JobBench margin is large enough to take seriously.
The warning: your cheapest frontier-agent option is now a metered API you don't control, from the vendor who used to guarantee you'd never be in that position. Plan for a world where that's the norm. Keep an abstraction layer between your agent and any single provider, so that when the next "first paid model" ships, switching is a config change and not a rewrite. The weights may keep flowing for last year's capability. The frontier is going behind the meter.
