What changed
Meta released Muse Spark 1.1 on July 9, its second model from Meta Superintelligence Labs since the original Muse Spark debuted in April. The model itself is not the news. The news is what shipped alongside it: the Meta Model API, a public preview that opens Meta's frontier model to outside developers for the first time on a paid, metered basis.
For a company that has spent years positioning Llama as the free, open-weight alternative to closed labs, this is a business model change, not just a model update. Muse Spark 1.1 is now available through meta.ai, the Meta AI app, and a developer API at api.meta.ai/v1, priced like any other commercial frontier model.
The pricing is the story
Meta priced the API at $1.25 per million input tokens and $4.25 per million output tokens, with $20 in free credits for new accounts before billing kicks in. Alexandr Wang, who leads Meta Superintelligence Labs, told CNBC the pricing was deliberately "aggressive and attractive" against Anthropic and OpenAI. Bloomberg reported Mark Zuckerberg pitched the same number differently: roughly a quarter of what comparable OpenAI and Anthropic offerings cost.
The API is also built for a frictionless switch. Meta made it compatible with both the OpenAI Chat Completions and Responses formats and Anthropic's Messages format, so a developer can generate a Meta API key and point existing integration code at Meta's endpoint without rewriting it. That is a direct pitch at teams already building on the two incumbents.
Who gets in first
Access is running through a public waitlist, but early partners are already live, including Replit, Cline and Box. Two of those, Replit and Cline, are agentic coding tools, which lines up with how Meta is positioning the model: as an engine for multi-agent orchestration rather than a general chatbot backend. Meta describes Muse Spark 1.1 as able to act as either a "primary agent" that plans and delegates work, or a "subagent" that executes a narrow task and escalates when it hits a wall.
That framing matters for who this affects first. Teams building agentic coding tools and internal automation now have a fourth serious paid API option alongside OpenAI, Anthropic and Google, at a price that undercuts all three on paper.
The benchmarks are a mixed picture
Muse Spark 1.1 does not lead on raw coding benchmarks. On SWE-Bench Pro it scored 61.5 against Claude Opus 4.8's 69.2, and on DeepSWE 1.1 it scored 53.3 against GPT-5.5's 67.0 and Opus 4.8's 59.0, according to Meta's published evaluation figures. Terminal-Bench 2.1 tells a similar story: 80.0 for Muse Spark 1.1 against 83.4 for GPT-5.5 and 82.7 for Opus 4.8.
Where it does lead is agentic tool use. On MCP Atlas, a benchmark for multi-step tool orchestration, Muse Spark 1.1 scored 88.1 against 82.2 for Opus 4.8 and 75.3 for GPT-5.5. On JobBench, a professional tool-use benchmark, it scored 54.7 against Opus 4.8's 48.4 and GPT-5.5's 38.3. Read together, the numbers describe a model that is not the strongest individual coder on the market but may be the strongest at coordinating other agents and tools, which is a different, and per Meta's own pitch more commercially relevant, capability for the agentic workloads enterprises are currently trying to scale.
Why it matters
Meta has built its AI identity around Llama being free and open. Muse Spark, first released in April under Wang following Meta's roughly $14 billion investment to bring him in from Scale AI, was already a break from that pattern as a closed model. The Meta Model API turns that break into a business line: Meta is now selling frontier-model access at a price low enough to pressure OpenAI and Anthropic's per-token economics, using compatibility with their own SDK formats to make switching cheap.
For developers already paying for agentic coding infrastructure, the practical effect is a lower-cost option that trails on raw coding accuracy but leads on the orchestration work agentic pipelines actually spend most of their compute doing. Whether that trade holds up outside benchmark tables is the question enterprise teams evaluating the waitlist will now have to answer for themselves.