Muse Spark: Meta AI’s Multimodal Reasoning Model for AI Coding and Agents

Meta Muse Spark is the AI model family behind the company’s agents. Meta Superintelligence Labs launched Muse Spark on April 8, 2026, and the company calls it a natively multimodal reasoning model with tool use, visual chain of thought and multi-agent orchestration, built for AI agents. If you searched for Muse AI, note that the name was also used by an unrelated video-hosting service, which says on its own page that it is now Skiv.

This guide covers what the model is, how it behaves as an agent, which versions exist, what they cost, where the limits are and how it compares with ChatGPT Work, Claude Cowork and xAI’s agent. Every vendor fact comes from the makers’ own pages, read on 22 October 2026, and any of it can change. RazzTools is not affiliated with any company named here.

Key takeaways

• The model launched in April and has had three updates, in July, August and September.
• A main agent can plan a hard task and hand parts of it to parallel subagents, which suits long tasks.
• Developer access costs $1.25 per million input tokens and $4.25 per million output tokens on the standard tier, as of October.
• The personal agent app, Muse, is free with a usage limit, with paid plans at $20 and $100 a month.

What is Muse Spark and which version is current?

The launch post calls it the first model in the Muse family, the first step on its scaling ladder and the first product of a ground-up overhaul of the company’s AI work. It reached the company’s app and website on launch day, with a private API preview for select developers. It also admitted gaps in long-horizon agentic systems and coding workflows. The post says the new training recipe reaches the same capability with over an order of magnitude less computing power than Llama 4 Maverick, the previous model, and for health answers it says physicians helped curate training data. Mark Zuckerberg’s launch-day Threads post added that future releases would include new open source models.

The developer documentation calls the latest version the one to use for new work. The family has grown since: Muse Wonder, a new model announced on October 12, is described as the most capable, and Muse Spark Open, released on October 15, ships downloadable weights and has 816 billion total parameters, 42 billion of them active. The documentation says to pick Wonder for long, complex tasks and Spark for high-volume or cost-sensitive work, such as subagent tasks.

How Muse Spark works as a coding and agent model

An agent plans, calls tools and keeps working over many steps, where a chatbot answers once. The company builds that behaviour in through training and tests it on agentic tasks and large workloads.

Parallel agents and subagents

At launch, the maker explained that the usual approach has one agent think for longer, while its approach adds more agents that collaborate, giving stronger results at comparable latency. Contemplating mode applies the idea and aims at the extreme reasoning modes of Gemini Deep Think and GPT Pro. For it, the company reports 58% on Humanity’s Last Exam and 38% on FrontierScience Research, and said it would roll out gradually.

The July model goes further. As the main agent it gathers context, makes a plan and delegates work to parallel subagents; as a subagent it sticks to its job and escalates back when needed. Muse Code, a terminal tool for agentic coding, can coordinate several persistent subagents on one task, and the documentation shows how to build coding agents on the model.

Computer use, coding benchmark results and long context

The company says 1.1 works across several applications, chooses between writing a script when automation is faster and clicking, adapts to new MCP servers and manages a context window of 1M tokens. In its own benchmark table for that model, the OSWorld-Verified score is 80.8 against 83.4 for Opus 4.8, and SWE-Bench Pro is 61.5 against 69.2 for that same rival. These are the maker’s numbers, not independent tests. For the latest version, it says the model asks clarifying questions when a prompt is ambiguous, confirms before consequential actions and, against the previous version, used about 20% fewer tool calls and 25% less token use in its engineers’ comparisons. It also says the model performs competitively with frontier models on several coding evaluations. The page for the latest version compares it with the newest models from OpenAI and Anthropic on benchmarks such as GDPVal-AA, OSWorld 2.0 and Terminal-Bench 2.1. Treat any one score as a data point, not a verdict.

Muse Spark pricing and who can use it

There are three routes: the consumer app, the Muse personal agent and developer access. The company’s API pricing page lists these standard pricing rates per million tokens, as of October.

The Contributor tier is cheaper because content may be used to improve the company’s products; the standard tier is not used that way. Reasoning tokens are billed as output. Meta released the Meta Model API as a public preview to US developers in July, and the August release brought expanded global access. The Help Center says Muse is free with a usage limit, with a Power plan at $20 a month and a Maximum plan at $100 a month. It adds that the app is in limited testing. The top model, Wonder Full, is for subscribers to the coding tool on its $50 and $200 plans.

Limits and risks of the model

• Availability: Muse is in limited testing and is not offered everywhere, so check your country first.
• Audio: the documentation says audio understanding in the latest version is not fully supported.
• Data: the cheaper tier lets the company train future models on your content.
• Agent actions: by default Muse stops for anything hard to undo, but an agent with access to email or payments can still err, so read each approval request.
• Test awareness: in April the company reported that Apollo Research saw the highest rate of evaluation awareness it had observed in a near-launch checkpoint, which makes test results harder to read.

How to get started: the consumer app, Muse and developer access

To try it as a user, open the consumer app or its website. For the agent, sign in with a Meta account in the Muse mobile app where it is offered. Developers create a key in the developer dashboard, point an existing SDK client at api.meta.ai/v1 and set the model to muse-spark-1.3, the newest version of Muse Spark. The Anthropic Messages format also works, so tools such as Claude Code can connect, and built-in web search returns cited answers. The launch post shows what the first release did.

Meta’s agents vs ChatGPT Work, Claude Cowork and other agents

These are not the same kind of product: Muse is a consumer agent, while the model is something developers build with. Here is how each maker’s pages describe the agents.

Other AI companies ship comparable agents. OpenAI’s help page says ChatGPT agent is no longer available and points to ChatGPT Work. The entry price and open weights are Meta’s edge, and a younger product and limited testing are its weak side. For Anthropic’s side, read our Claude plans guide.

Who Muse Spark is for, and who should skip it

It suits developers who want a cheap, drop-in compatible model with subagent support, people curious about a free personal agent and teams that want open weights. It also suits anyone comparing AI models before choosing one. Skip it if you need audio understanding on the newest version, if Muse is not offered where you live, or if you need a vendor with a long enterprise record.

RazzTools has no product that gives access to these models. It lists other AI tools in its tools catalog, with prices in USD or BDT and support every day from 8 AM to 3 AM Bangladesh time. If you want a different AI assistant to use meanwhile, the Perplexity group buy is one option, and it is a separate product from anything offered by the makers above.

FAQ

What is Muse Spark?

It opens the Muse line of models, launched in April 2026 for agentic work and reasoning over text, images and more.

Is it free to use?

Muse has a free tier with a usage limit, where it is offered. Developer access is pay as you go from $1.25 per million input tokens.

What is Muse AI, and is it the same thing?

Muse AI now most often points to the company’s agent and models. Spark is the model family, and Muse is the agent app.

Is there a Muse Spark 1.1 API?

Yes. The preview opened in July with the model ID muse-spark-1.1, and later versions were added.

Is the model open source?

The original releases are closed. An open-weight release came on October 15 under the Muse Model License, which has limited commercial terms.