Muse Spark 1.1: 1M context aimed at agentic coding

Meta has released Muse Spark 1.1 as its strongest agentic and coding model yet, and paired the launch with the first public preview of the new Meta Model API. In Meta's own announcement, the company describes Muse Spark 1.1 as a multimodal reasoning model with major gains in tool use, computer use, coding, and multimodal understanding — exactly the cluster of capabilities that matter when a model has to execute work instead of only chat about it.

Definition: Muse Spark 1.1 is Meta's new multimodal reasoning model for agentic tasks, coding, computer use, and long-running tool-based workflows.

Example: Meta says Muse Spark 1.1 can plan work, orchestrate parallel subagents, use tools, manage a 1 million token context window, and adapt across real coding and computer-use harnesses.

Key takeaway: This is not just a model refresh inside Meta AI; it is Meta's clearest attempt yet to sell a frontier agent-execution layer to outside developers.

Business impact: Teams building coding agents or tool-using assistants should test Muse Spark 1.1 on real workloads, because the launch combines strong agentic claims with unusually aggressive API pricing.

What did Meta actually launch?

Meta launched two things together: Muse Spark 1.1 itself, and a public preview of the Meta Model API that exposes the model to developers. That matters because the original Muse Spark rollout was mostly limited to internal use and selected partners, while this release turns the model into something the wider market can actually evaluate. Meta also says Muse Spark 1.1 is available now in "Thinking" mode inside Meta AI, which means this is both a product-model update and a platform-access update at the same time. See also Meta Says Watermelon Is Catching Up to GPT-5.5. What Is Actually Known?. Background: Meta's Glimmer targets always-on agents without the cloud. More on this: Muse Spark 1.1: Meta’s Cybersecurity Test Explained.

The timing matters too. Meta shipped Muse Image earlier this week, and Muse Spark 1.1 now extends the Muse line into a more explicit developer and agent-builder offering. Read together, the two launches look less like isolated model drops and more like Meta trying to make the Muse family a full stack for creators, apps, and agentic workflows.

Why is Muse Spark 1.1 being framed as an agentic model, not just a coding model?

Meta's own framing is broader than coding alone. In the launch post, Meta says Muse Spark 1.1 is built for agentic tasks, with major improvements in planning, orchestration, tool use, computer use, and multimodal understanding. The company says the model can zero-shot generalize to new native tools, MCP servers, and custom skills, and can work either as a main agent that gathers context and delegates across subagents or as a subagent that stays within a narrower job.

That distinction matters because a strong coding model is not automatically a strong agent model. A coding model can write code well and still fail at managing context, deciding when to call a tool, recovering from errors, or escalating correctly. Muse Spark 1.1 is interesting precisely because Meta is claiming improvements in the full execution loop rather than only in raw code generation.

What are the most important technical claims in the launch?

The biggest practical claim is the 1 million token context window. Meta says Muse Spark 1.1 can actively manage that context, remember earlier actions, retrieve information from much earlier work, and compact its context in a way that preserves important steps for later tasks. If that holds up in real use, it makes the model more relevant for long-running coding sessions, research workflows, and tool-heavy tasks where context decay is a real failure mode.

The second important claim is around computer use and orchestration. Meta says Muse Spark 1.1 is trained to decide when to automate with scripts and when to interact directly with an interface, instead of reasoning through every desktop click the same way. That is a more mature agentic claim than simply saying the model can use a browser, because it suggests Meta is optimizing for end-to-end workflow speed and reliability rather than demo-style interaction.

How much does Muse Spark 1.1 cost?

Pricing may be the most strategically important part of the announcement. CNBC reports that Meta will charge $1.25 per million input tokens and $4.25 per million output tokens for Muse Spark 1.1, with $20 in free credits for every new API account. Alexandr Wang described that pricing to CNBC as "very aggressive and attractive" compared with Anthropic and OpenAI.

That price matters because agentic workloads can consume far more tokens than a one-shot chat interaction. Planning steps, tool results, retries, traces, and self-checks all add up. A lower model price can change which agent architectures are financially viable, especially for coding assistants and long-running automated tasks where cost compounds fast.

Where does Muse Spark 1.1 look strongest?

Meta is clearly pushing three use cases hardest: agentic task execution, coding, and computer use. In the launch post, the company says Muse Spark 1.1 performs well on complex coding tasks in large codebases, can implement features and code migrations, and supports common agentic coding patterns such as planning mode, goal conditioning, subagent delegation, and context compaction. Meta also highlights multimodal use cases where perception and action happen together, such as creating listings from smartphone-shot video.

The signal for businesses is that Meta is not pitching Muse Spark 1.1 as a general-purpose chatbot first. It is pitching Muse Spark 1.1 as a model that can act through software, tools, and interfaces — much closer to the role described in what an AI agent actually is than to a simple assistant experience. Background: Gemini Spark rolling out to Google AI Pro users in the US.

What does the safety material add to the story?

Meta published a separate Muse Spark 1.1 evaluation report alongside the launch, which is notable in itself because it treats the API release as a higher-affordance deployment surface than consumer chat. According to Meta's summary, the company evaluated the model under its Advanced AI Scaling Framework and says the deployed version operates within acceptable risk bounds after mitigations.

That does not mean agent builders should treat the model as plug-and-play safe. It means Meta is explicitly acknowledging that a model with developer-controlled prompts, tool use, and agentic scaffolding creates a different risk profile from a plain chat interface. For teams considering Muse Spark 1.1, the right read is not "safe therefore done" but "more capable therefore worth stricter testing."

How should teams benchmark Muse Spark 1.1 now?

The first useful test is not a synthetic leaderboard. It is a fixed workload from your actual environment: a codebase task, a support workflow, a browser flow, or a research loop that already breaks weaker models. Muse Spark 1.1 looks most interesting where three things matter together: long context, tool use, and cost. That makes it a natural candidate for coding agents, internal ops assistants, and multi-step task runners that spend too much on more expensive frontier models today.

A good benchmark should track more than final answer quality. It should log tool choices, retries, latency, context growth, review burden, and whether the model finishes the task cleanly or only gets most of the way there. That is especially important now, because Meta is making a cost-performance pitch, not only an intelligence pitch.

What should businesses watch next?

The next real signal is whether independent developers reproduce Meta's claims in production-like coding and agentic setups. Meta says Muse Spark 1.1 works well with popular harnesses and can orchestrate long-running tasks across tools and subagents. If outside teams confirm that at the advertised price, Muse Spark 1.1 could become relevant not because it tops every benchmark, but because it lowers the cost of running serious agent workflows at scale.

The more strategic question is what this does to the frontier-model pricing conversation. If Meta is willing to sell an explicitly agentic, coding-oriented model this cheaply, the market shifts from "which lab has the smartest model" to "which model is good enough to run constantly." That is a much more operational question — and a more important one for teams building real agents.

Want help comparing new frontier models on your own workflows before you switch? Get in touch.