Meta opened its model API on July 9. Not the Llama weights download — the paid commercial one, the kind where you get an invoice and a rate card and a bill at the end of the month. First time for Meta. Worth noting.

The model is Muse Spark 1.1, second release from Meta Superintelligence Labs. Artificial Analysis puts it at 51 on their Intelligence Index — up 8 points from Muse Spark 1.0's 43 in April, which is a fast gain, but 51 still places it in solid mid-tier territory. Not frontier. Priced at $1.25/$4.25 per million tokens.

If you stop there, the story is short: mid-tier intelligence at a mid-tier price, fine, next. Don't stop there.

MCP Atlas is the scaled tool-use benchmark that tests models across coordinated multi-tool agent environments. Muse Spark 1.1 scores 88.1. Opus 4.8 and GPT-5.5 both sit in the high 70s to low 80s on the same test. JobBench — professional-context tool use across realistic work tasks — Muse Spark 1.1 scores 54.7 against Opus 4.8's 48.4 and GPT-5.5's 38.3.

That GPT-5.5 number deserves a second read. 38.3 on JobBench. Muse Spark 1.1 at 54.7. The gap is not close.

Source spread

• Meta AI — Introducing Muse Spark 1.1 — [hype]. First-party launch post. Covers context management, parallel subagents, API surface, and benchmark highlights. Leads with the tool-use numbers, not the overall Intelligence Index.
• Artificial Analysis — Muse Spark 1.1 Intelligence Index — [builder]. Independent evaluation with cross-model Intelligence Index comparisons; source of the 51 and 43 figures.
• MarkTechPost — Meta Superintelligence Labs releases Muse Spark 1.1 — [builder]. Technical breakdown of API surface, MCP Atlas and JobBench benchmark methodology, and pricing.
• AI Weekly — Meta prices Muse Spark 1.1 API — [skeptic]. Notes the aggressive pricing but questions whether mid-tier intelligence limits the viable use case.

Pros & cons

What's real:

• MCP Atlas 88.1 is a genuine lead in the category that matters most for agent builders right now. If you are building products where models coordinate across many tools in parallel, that benchmark is telling you something real about how the model handles the specific kind of work you care about.
• 1M token context with active compaction means the model manages its own context window — automatically compacting long sessions while preserving the critical decision chain. You shouldn't have to write context management logic around this. Worth verifying, but the design is right.
• Parallel subagents as a first-class API feature, not a prompt engineering workaround. For multi-step orchestration, this matters in ways that are hard to replicate with single-threaded reasoning.
• $1.25/$4.25 per million tokens plus $20 in free credits makes evaluation cost essentially zero. That's an unusually low-friction on-ramp for a frontier-adjacent model.

What deserves a side-eye:

• Intelligence Index 51 is mid-tier. For knowledge synthesis, reasoning depth, or open-ended code generation, Muse Spark 1.1 is not in the conversation with Fable 5 or GPT-5.5 Sol. The tool-use lead doesn't transfer across all task types, and it would be a mistake to treat it as a general-purpose replacement.
• "High 70s to low 80s" for Opus 4.8 and GPT-5.5 on MCP Atlas is imprecise enough that the gap might be smaller than it looks. Both Anthropic and OpenAI know what benchmark they fell behind on this week. I'd give it two quarters before that lead shrinks.
• US-only public preview. If your team or infrastructure is outside the US, this is a near-term wait with no published timeline.
• First-generation paid API from a company whose primary business is advertising. Meta has good reasons to keep Muse Spark priced aggressively right now. That doesn't mean the pricing is permanent.

Samwise's take

What builders need to know

• Route MCP-heavy workloads first. The 88.1 MCP Atlas score is the specific claim to verify in your environment. Run your actual tool-call sequences against Muse Spark 1.1 before assuming the benchmark generalizes to your stack.
• Use the $20 free credits for evaluation. That's enough tokens to run a meaningful comparative agent test. Don't commit budget until you've verified the tool-use lead holds for your specific tools.
• Mid-tier planning, strong tool execution. Muse Spark 1.1 is not a replacement for frontier models on reasoning tasks. Think of it as the tool-execution layer in a multi-model stack, not the planning layer.
• Active context compaction at 1M tokens is the feature to test if you've been managing rolling context windows manually. Let the model handle it and measure whether the critical reasoning chain survives the compaction.
• US-only for now. Check the availability page before planning any production deployment outside the US. No expansion timeline has been published yet.

Further reading

• Meta AI — Introducing Muse Spark 1.1 — official launch post
• Meta — Muse Spark 1.1 evaluation report — full benchmark methodology
• Artificial Analysis — Muse Spark 1.1 Intelligence Index — independent cross-model evaluation
• MarkTechPost — Meta Superintelligence Labs releases Muse Spark 1.1 — technical coverage and MCP Atlas details
• AI Weekly — Meta prices Muse Spark 1.1 API — pricing comparison to Anthropic and OpenAI mid-tier