On July 9, 2026, Meta Superintelligence Labs shipped Muse Spark 1.1 — its second model after April's original Muse Spark — and did something Meta had never done before: it started charging developers to use its own model, via the new Meta Model API. For a company synonymous with open weights, that's the real headline.
This is a developer-focused look at what Muse Spark 1.1 offers, where it fits, and how to evaluate it against GPT-5.6 and Grok 4.5. If you're deciding whether to build on it, the interesting parts are the 1-million-token context, the agentic feature set, and the strings attached to a preview API.
What it is
Muse Spark 1.1 is a multimodal reasoning model built explicitly for agentic tasks. Early partners describe it as a complete agentic foundation — long-context handling paired with coding and reasoning aimed at large-scale agent workloads. The defining spec is its context window: 1 million tokens, enough to feed an entire codebase, a full legal contract set, or months of operational logs in a single call.
1M-token context, used well
A million-token window changes which problems are tractable without a retrieval pipeline:
- Reason over a whole repository when a change spans many files and cross-references matter.
- Analyze a complete document set — contracts, filings, specs — in one pass instead of chunk-and-stitch.
- Feed long operational histories (logs, incident timelines) for root-cause analysis.
The usual caution applies: a bigger window is not license to dump everything in. Long inputs cost more and can bury the signal. Treat 1M tokens as headroom for genuinely large context, not a reason to skip relevance filtering.
Agentic features
Beyond context, Muse Spark 1.1 is positioned for autonomous work: reports describe computer use (operating desktop/browser/mobile interfaces) and parallel subagent delegation — the same direction OpenAI and Anthropic have pushed with their agent stacks. As always with computer-use agents, the safety posture matters: scope access, keep a human on consequential steps, and test where a mistaken action is recoverable.
The Meta Model API
Muse Spark 1.1 debuts on the Meta Model API in public preview. Pricing is $1.25 per million input tokens and $4.25 per million output — among the more aggressive frontier prices this month — with usage-based billing after the preview. The catch: access is US-only at launch. Developers outside the US can't build on it yet.
Why "Meta charging" matters
Meta's brand has been open weights you download and run yourself. Shipping a frontier model only behind a paid API is a strategic pivot: it puts Meta in direct commercial competition with OpenAI and Anthropic for developer spend, and it signals that Meta's most capable models may increasingly live behind an API rather than a download link. For teams, that means Muse Spark 1.1 is a vendor relationship, not a self-hosted asset — plan for rate limits, pricing changes, and regional gating accordingly.
How to evaluate it for your stack
- Region: US-only preview — confirm access before designing around it.
- Context need: the 1M window is the standout reason to choose it; if your workloads fit in 200K, the advantage shrinks.
- Agentic fit: strong candidate for long-horizon, tool-using agents — benchmark it on your actual agent loop, not a chat prompt.
- Compare honestly: put it head-to-head with GPT-5.6 (tiered reasoning) and Grok 4.5 (token efficiency) on your eval set — see the July 2026 showdown.
- Benchmarks: treat Meta's launch claims as directional and verify independently before switching production traffic.
Bottom line
Muse Spark 1.1 gives developers a genuinely large-context, agent-native model at an aggressive price — and it marks the moment Meta joined the paid-API race. If you need million-token context and you're in the US, it's worth a serious evaluation. If you're outside the US or your context fits comfortably elsewhere, keep it on the watch list and revisit when the preview opens up.