IMPLICATOR
Meta Superintelligence Labs released Muse Spark 1.1 on Thursday and opened a public preview of the Meta Model API, the first time the company has charged outside developers for access to a model it built. Developers pay $4.25 per million output tokens and $1.25 per million input tokens, chief AI officer Alexandr Wang told CNBC. Those rates come to roughly 25% of the cost advertised by other top models from OpenAI and Anthropic, Bloomberg reported.
Key Takeaways
- Meta opened a public preview of the Meta Model API on Thursday, charging outside developers for one of its own models for the first time, at $1.25 per million input tokens and $4.25 per million output tokens.
- Bloomberg reported the rates come to roughly 25% of what OpenAI and Anthropic advertise for top models; Reuters noted they still sit above OpenAI's GPT-5 mini and Anthropic's Claude Haiku 4.5.
- On Meta's own launch table, Muse Spark 1.1 leads four of twelve benchmarks, all in tool use, while Claude Opus 4.8 leads five and GPT-5.5 leads three. It trails Opus 4.8 on SWE-Bench Pro coding, 61.5 to 69.2.
- The interface arrived five weeks after a Meta spokesman told Reuters on June 3 that it would ship that month.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
$1.25 in, $4.25 out
Every new account starts with $20 in credits before the meter runs, Wang said. Cached input costs $0.15 per million tokens, and the web-search grounding tool runs $2.50 per 1,000 queries, according to The Decoder. Meta's API documentation lists the context window at 1,048,576 tokens, and the interface is compatible with OpenAI's SDK, with structured output, parallel tool calling and prompt caching.
The public preview is open only to developers in the United States, The Verge reported. New users join a waitlist and some early partners already have access, a Meta spokesperson told CNBC. Meta is serving the model from its own properties for now rather than listing it on third-party marketplaces such as OpenRouter.
"Since this is not an open source model, this is I think the first time that we're doing a real serious API," Chief Executive Mark Zuckerberg said in a Bloomberg interview before the release. "And the pricing is going to be very aggressive and attractive." The pricing from some of the other labs, he said, "is very extreme and has very high margins." Reuters reported the Muse Spark rate sits above OpenAI's entry-level GPT-5 mini and Anthropic's low-cost Claude Haiku 4.5, and below Anthropic's higher-end Claude Sonnet 4.6. The Decoder put Opus 4.8, GPT-5.5 and Fable 5 between $25 and $50 per million output tokens.
Meta leads four of the twelve benchmarks it chose
Meta published its own launch table, with rivals shown in their strongest reasoning modes. Across the twelve benchmarks Meta selected, Muse Spark 1.1 leads four, Anthropic's Claude Opus 4.8 leads five and OpenAI's GPT-5.5 leads three, per The Decoder's reading of the table. The leads are in tool use. Muse Spark 1.1 scored 88.1 on MCP Atlas, ahead of Opus 4.8 (82.2), Google's Gemini 3.1 Pro (78.2) and GPT-5.5 (75.3), and 54.7 on the professional tool-use test JobBench, where Gemini 3.1 Pro managed 15.9. On the coding tests the launch is named for, it trails: 61.5 on SWE-Bench Pro against Opus 4.8's 69.2, and 53.3 on the long-horizon DeepSWE 1.1 against GPT-5.5's 67.0; Opus 4.8 also takes computer use on OSWorld-Verified, 83.4 to 80.8. Meta trained the model for coding "in service of overall agentic capabilities," Wang told CNBC, describing agents that autonomously carry out multistep tasks. Meta did not say how Muse Spark 1.1 compares with the most recent models from OpenAI and Anthropic, Fortune reported, and on one open-source coding leaderboard it still lagged Anthropic's Mythos 5 and Fable 5 and OpenAI's GPT-5.6. The table is Meta's own, and the company has fielded questions about its numbers before: in April 2025, a former Meta AI executive said on X that claims Meta trained on test sets were "simply not true and we would never do that."
Meta's 2026 capex guidance runs to $145 billion
Meta shares rose nearly 2% on Thursday. Shay Boloor, chief market strategist at Futurum Equities, told Reuters: "If Muse Spark 1.1 is genuinely competitive with Claude and Codex on coding, then Meta may finally have a much clearer monetization bridge from AI models to paid developer tools."
Get Implicator.ai in your inbox
Strategic AI news from San Francisco. No hype, no "AI will change everything" throat clearing. Just what moved, who won, and why it matters. Daily at 6am PST.
No spam. Unsubscribe anytime.
Joe Tigay, a portfolio manager at Equity Armor Investments, read the discount as a concession. "They are forced to compete on price and raw computing real estate rather than model superiority," he told Business Insider. Nick Jones, a senior analyst at BNP Paribas Equity Research, wrote to investors on Thursday that Meta is "well positioned to generate ample revenue to support its spending," citing fees for external use of its models alongside advertising share gains and subscription revenue, according to a note obtained by Business Insider. Meta lifted its 2026 capital-expenditure guidance to between $125 billion and $145 billion, up from $115 billion to $135 billion, and advertising still accounts for about 98% of company revenue, according to its first-quarter results.
The API Meta said would ship in June
Meta shipped the interface five weeks late. A company spokesman told Reuters on June 3 that Meta was testing the API with early partners and looked forward to releasing it that month, after the Wall Street Journal reported repeated delays and no scheduled launch date. On July 2, Zuckerberg told an internal town hall that Meta's restructuring bets "haven't come to fruition yet." Wang posted on X the same day that a Muse Spark update with gains in coding and agent tasks was coming.
Know someone who'd find this useful? ✉️ Email it to a friend in one click, or they can subscribe free here.
Zuckerberg announced Thursday's release on X himself, calling Muse Spark 1.1 "a strong agentic and coding model at a very low price." It was his first post on the platform in three years, TechCrunch reported; the last came in July 2023.
Cline chief executive Saoud Rizwan, whose company was given access before the preview opened, said Meta is "clearly building for serious agentic coding" at a price "that makes it viable to run real coding workloads at scale."
Wang said Meta remains "committed to open source" and that a variant of Muse Spark is in development that the company intends to release openly, declining to say when. Muse Image, the image model Meta shipped on Tuesday, is not yet available through the API.
Frequently Asked Questions
What does the Meta Model API cost?
Developers pay $1.25 per million input tokens and $4.25 per million output tokens, chief AI officer Alexandr Wang told CNBC. Cached input costs $0.15 per million tokens, and the web-search grounding tool runs $2.50 per 1,000 queries. Every new account starts with $20 in free credits before pay-as-you-go billing begins.
Is Muse Spark 1.1 actually better at coding than its rivals?
Not on the coding tests in Meta's own launch table. Muse Spark 1.1 scored 61.5 on SWE-Bench Pro against Claude Opus 4.8's 69.2, and 53.3 on DeepSWE 1.1 against GPT-5.5's 67.0. Its leads are in tool use, where it scored 88.1 on MCP Atlas. Wang said Meta trained coding capability in service of agentic performance.
Can developers outside the United States use it?
Not yet. The public preview is open only to developers in the United States, The Verge reported. New users join a waitlist. Meta is also serving the model from its own properties rather than listing it on third-party marketplaces such as OpenRouter.
Is Muse Spark 1.1 open source?
No. It ships without open weights, the second Meta model to do so after the original Muse Spark. Wang said Meta remains committed to open source and that a variant of Muse Spark is in development that the company intends to release openly, but he declined to say when.
Why does a paid API matter for Meta?
It is a new revenue stream from a model business that previously charged nothing. Meta lifted its 2026 capital-expenditure guidance to between $125 billion and $145 billion, and advertising still accounts for about 98% of company revenue, so investors have pressed for evidence the AI spending produces sellable products.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
Marcus Schuler
Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai
