Meta Ships Muse Spark 1.1 and Opens Its First Paid Model API
Meta Superintelligence Labs' closed-weight agent model and $1.25/$4.25 API mark the company's definitive pivot from open-weight goodwill to metered AI revenue.
Meta launched Muse Spark 1.1 on July 9, 2026, via the new Meta Model API — the company's first paid developer API — priced at $1.25 per million input tokens and $4.25 per million output, in a US-only public preview behind a waitlist with $20 in free credits, according to coverage of the Meta Model API documentation.
Context
Meta built its AI reputation on open-weight Llama releases. Muse Spark 1.1, from Meta Superintelligence Labs (MSL) under Alexandr Wang, is closed-weight — and Zuckerberg said so himself in his first X post in roughly three years: "Since this is not an open source model, this is I think the first time that we're doing a real serious API," quoted by MLQ.ai.
What changed
- First paid Meta API. Public preview for US developers only at launch; $20 free credits; early partners include Replit, Cline and Box.
- Model profile: multimodal input, text output, 1M-token context, agentic focus — tool use, computer use across desktop/browser/mobile, parallel subagents, and native MCP support.
- Consumer access: free inside the Meta AI app's "Thinking" mode.
Why it matters
This is Meta's definitive business-model pivot: converting AI from an open-ecosystem goodwill play into a direct revenue line, and buying developer share on price. At $1.25/$4.25, Spark 1.1 undercuts GPT-5.6 Sol ($5/$30) and roughly matches Luna ($1/$6), making it the cheapest serious agent model of the July wave. It lands as the open-weight center of gravity shifts to Chinese labs and Thinking Machines — Meta is now competing with the closed frontier on API economics rather than leading the open one.
Details
Independent measurement by Artificial Analysis puts Spark 1.1 at 51 on its Intelligence Index (xhigh effort), up from 43 for its predecessor. Meta claims the #1 spot on the MCP Atlas agentic benchmark (88.1) — a Meta-measured result — while the model trails Claude Opus 4.8 by roughly seven points on SWE-bench Pro, per eCorpIT's evaluation coverage.
Limitations and caveats
Meta's headline benchmark wins are vendor-run; coding performance trails the frontier. The API's US-only gating and waitlist limit immediate availability, and no international expansion date has been published.
Sources
- The Planet Tools — Gemini 3.5 Flash vs Muse Spark 1.1 (pricing and availability) (aggregator, cites Meta Model API docs)
- MLQ.ai — Meta launches Muse Spark 1.1 API (aggregator)
- eCorpIT — Muse Spark 1.1 coding-agent evaluation (aggregator)
*Update note: This post was last reviewed on 2026-07-22. Watch for international API expansion and independent replications of Meta's MCP Atlas claim.*
Sources
- The Planet Tools — Gemini 3.5 Flash vs Muse Spark 1.1 — aggregator
- MLQ.ai — Meta launches Muse Spark 1.1 API — aggregator
- eCorpIT — Muse Spark 1.1 coding-agent model evaluation — aggregator
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.