Meta released Muse Spark 1.1 on July 9, 2026. It's their first closed-weight model behind a paid API, and the benchmarks tell an interesting story: it beats Opus 4.8 and GPT-5.5 on tool-use and agent tasks, trails them both on pure coding. That split matters for how you route it.
Meta Just Went Closed (Sort Of)
For years, Meta's play was Llama. Open weights, self-hostable, no API lock-in. That playbook built a huge developer ecosystem and kept Anthropic and OpenAI honest on pricing. Now Meta has a paid closed model at $1.25 per million input tokens and $4.25 per million output tokens.
To put that in context: Opus 4.8 runs around $15/$75. GPT-5.5 Sol is $5/$30. Muse Spark 1.1 at $4.25 output is competing on price with Flash-tier models while claiming frontier-tier performance on agent benchmarks. That's a meaningful price point if the benchmarks hold.
They're not abandoning Llama. Muse Spark is a separate product line from Meta Superintelligence Labs. Open weights for the community, closed model for the API business. It's the same move Google made with Gemini vs. Gemma, or the way Mistral keeps Mixtral open while selling Mistral Large. What's new is Meta doing it, because they spent years specifically not doing it.
Where It Wins
The MCP Atlas benchmark measures scaled tool use: multi-step orchestration, parallel tool calls, agent loops. Muse Spark 1.1 scored 88.1 there. Opus 4.8 hit 82.2. GPT-5.5 was at 75.3.
JobBench is similar: professional tool-use scenarios. Muse Spark: 54.7. Opus 4.8: 48.4. GPT-5.5: 38.3. Those are not marginal gaps. The model was clearly built for agentic workflows, not just trained on standard benchmarks and hoping for the best.
Finance Agent v2 and HealthBench Professional also go to Muse Spark. When you're building agents that call external APIs, chain tool calls, or need to maintain context across long multi-step runs, the 1M token context window plus these results make a reasonable case for trying it.
Where It Trails
Don't route it for code generation. SWE-Bench Pro: Opus 4.8 hits 69.2 to Muse Spark's 61.5. On DeepSWE 1.1, it's 53.3 vs GPT-5.5's 67.0 and Opus 4.8's 59.0. Terminal-Bench: both GPT-5.5 and Opus 4.8 beat it.
These aren't close. If your agent writes code, reviews PRs, debugs errors, or generates structured output that has to be syntactically correct, Muse Spark is not the right call yet. The tool-use wins don't carry over to the coding layer.
I haven't run it against my own codebases, so I'm going off the benchmark data. But the split is consistent enough across multiple independent evals that I'd treat it as real until I see otherwise.
How I'd Route It
This tells you something concrete. If you're building an orchestrator that decides what to do, calls tools, manages multi-step workflows, and handles long context, Muse Spark 1.1 at $1.25/$4.25 is worth evaluating seriously. You're paying a fraction of Opus 4.8 and probably matching or beating it on the metrics that matter most for orchestration.
If you also need code generation at the same step, keep Opus or GPT-5.5 there. The price gap is large enough that a hybrid routing setup makes sense: Muse Spark for orchestration and tool dispatch, stronger model only where you need real code or complex structured reasoning.
I'd set up a task-completion rate eval over a week of production traffic before committing. The benchmarks are a prior, not a guarantee.
The Llama Question
If you're running self-hosted Llama for cost reasons, this doesn't change that. Meta isn't pulling Llama. What changes is the option set: you now have a hosted Meta API priced to compete with Flash-tier models while claiming frontier-tier agent performance. The question isn't "do I switch from Llama" but "is there a hosted Meta model worth the per-token cost for the parts of my stack that need API-grade reliability."
For teams that were using Llama 4 Scout or Maverick specifically because there was no good closed Meta option, that gap is now filled. Whether Muse Spark 1.1 is better depends entirely on what you're building.
Mark Zuckerberg returned to X after three years to announce this. I'm still not sure what to make of that, but the model is real and the pricing is competitive.