For most of the last three years, Meta was the odd one out in the AI race. OpenAI, Anthropic, and Google competed on closed frontier models, while Meta gave its Llama models away as open weights. In 2026 that changed. Meta Superintelligence Labs launched a new model family called Muse Spark in April, shipped four updates by September, and then launched a consumer AI agent that is simply called Muse.
If you run a business in Canada, the name is now showing up in two places at once: in headlines about benchmark races, and in the Meta apps your customers already use every day. This post explains what Muse actually is, how it stacks up against ChatGPT, Claude, and Gemini, what it costs, and the privacy questions worth asking before anyone on your team starts using it for work.
Key Takeaway
"Muse" now means two different things. Muse Spark is Meta's model family, the engine. Muse is Meta's personal AI agent, a product built on top of it. Most of the confusion in coverage comes from mixing the two.
Muse Spark: Meta's new model family
Meta announced Muse Spark on April 8, 2026, describing it as the company's most powerful model yet and the first in a new family built by Meta Superintelligence Labs. It is a multimodal reasoning model, meaning it works with text and images and can think through a problem before answering. At launch it came with an Instant mode, a Thinking mode, and a Contemplating mode that runs several reasoning agents in parallel on harder problems.
It launched inside meta.ai and the Meta AI app, and began rolling out across WhatsApp, Instagram, Facebook, Messenger, and Meta's AI glasses. The launch was US-first, and developer access through an API started as a private preview only.
Source: Meta Newsroom, Introducing Muse Spark (April 2026)
The bigger story is the shift in strategy. Unlike Llama, the Muse Spark models are closed. Meta said at launch that it hoped to open-source future versions, and in August it released a smaller open-weight model, Muse Glimmer, under the Apache 2.0 licence. As of this writing, though, the flagship Muse Spark models themselves have not been released as open weights.
Source: Meta AI Research, Introducing Muse Glimmer (August 2026)
How fast it has moved
Meta has shipped at a pace that is unusual even by 2026 standards:
- Muse Spark 1.1 (July 9): better at agentic, multi-step work, with a one-million-token context window and a public preview of the Meta Model API.
- Muse Spark 1.2 and Muse Code (August 5): Muse Code is a coding agent that runs in the terminal on macOS and Linux.
- Muse Glimmer (August 10): a 30-billion-parameter open-weight model small enough to run on a single consumer graphics card once quantized.
- Muse Spark 1.3 (September 2): stronger on long, multi-step tasks, and better at asking clarifying questions before it acts.
Source: Meta AI Research, Introducing Muse Spark 1.3 (September 2026)
How it compares with ChatGPT, Claude, and Gemini
Vendor benchmarks are marketing, so the more useful numbers come from independent testers. Artificial Analysis, which runs the same evaluation suite across models, scored the original Muse Spark at 52 on its Intelligence Index in April, fourth behind Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6. Its clear weakness was agentic work: long tasks that require planning, tool use, and staying on track.
By September the picture had changed. In its review of Muse Spark 1.3 on September 2, Artificial Analysis scored the top configuration at 62, behind only two Anthropic models at the time, and the next configuration down tied OpenAI's GPT-5.6 Sol. The same review flagged regressions in long-context reasoning and factual accuracy, and VentureBeat noted that the best-scoring version was only available to a limited set of partners. Treat these as snapshots: Artificial Analysis has since updated its index methodology, and both OpenAI and Anthropic released new models later in September.
Source: Artificial Analysis, Muse Spark 1.3 review (September 2026)
Source: VentureBeat on Muse Spark 1.3 availability (September 2026)
What Muse Spark is good at Competitive reasoning scores in September 2026 Native to WhatsApp, Instagram, and Facebook Low API prices, especially on the contributor tier Fast release cadence
Where to be cautious Best-scoring version not broadly available Flagged regressions in long-context accuracy Flagship models are closed, not open weights Fewer business admin and data controls than mature rivals
The honest summary for a business owner: Muse Spark is now a genuine frontier model, not an also-ran. That does not automatically make it the right tool for your team. For most Canadian small businesses, the decision still turns on data controls, integrations with the tools you already use, and how well your team knows the product, not on a few points of benchmark score. Our ChatGPT, Claude, and Gemini comparison covers those practical factors.
Muse, the personal agent
On September 8, Meta launched Muse as a personal AI agent: an assistant that can take actions on the web on your behalf, not just answer questions. According to Meta, each person's agent runs in its own secure virtual machine, and a separate oversight agent approves actions before Muse takes them on the internet. It is available through the Muse app, the web, and WhatsApp. TechCrunch reported a free tier plus paid plans at US$20 and US$100 a month.
Source: Meta Newsroom, Introducing Muse (September 2026)
Source: TechCrunch on Muse pricing and trust (September 2026)
The launch was US-first. Later in September, Canadian outlet iPhone in Canada reported, citing Meta, that Muse was available in Canada on iOS and the web, with Android to follow. At Meta's Connect event on September 23, the company added connectors for retailers and for Shopify and Stripe, and said it expects to earn a small fee on transactions over time. That last detail matters: an agent that buys things for people is, eventually, a new storefront.
Source: iPhone in Canada, Muse available in Canada (September 2026)
The privacy question
This is where a business owner should slow down. Since December 16, 2025, Meta has used people's interactions with Meta AI to personalize the content and ads they see, in most regions, with some sensitive topics such as health, politics, and religion excluded. Meta's announcement did not list Canada among the excluded regions. Separately, Meta says the Muse agent does not share a person's conversations or the data in their virtual machine with its ad systems, and that users can opt out of having their data used for training.
Source: Meta Newsroom, using AI interactions to improve recommendations (October 2025)
Two practical points follow. First, the Meta AI assistant inside WhatsApp or Instagram and the Muse agent are not governed by identical data rules, so staff should not assume one policy covers both. Second, nobody on your team should paste client names, contracts, health information, or financial details into any consumer AI tool, Meta's or anyone else's, without written guidelines on what is allowed. That is the same advice we give about free ChatGPT or Gemini accounts. It is not legal advice; for regulated work, your privacy officer or counsel should review the guidelines before you adopt them.
What it means for your business
Most Canadian businesses do not need to do anything about Muse this month. But three groups should pay attention:
- Businesses that sell through WhatsApp, Instagram, or Messenger. Meta's Business Agent, launched globally in June, can answer customer questions, recommend products, book appointments, and qualify leads inside those apps, and was free to activate at launch with paid tiers promised later. If your customers already message you there, it is worth a structured test. Meta has not said which model powers it, so do not assume it is Muse Spark.
- Retailers and ecommerce brands. The Muse agent's shopping connectors are an early signal that AI agents will increasingly research and buy on a customer's behalf. Clear product data, accurate stock and pricing, and good reviews matter more when the shopper is an agent. This is the same groundwork as generative engine optimization.
- Teams building their own AI tools. The Meta Model API is priced aggressively, and the cheapest contributor tier trades lower prices for letting Meta use your prompts for training. That is a poor trade for anything involving client data, and a reasonable one for public or synthetic content.
Source: Meta Newsroom, Meta Business Agent (June 2026)
Our view
Muse is worth knowing about, not worth rebuilding around. The model is now competitive, the agent is interesting, and Meta's distribution through apps your customers already use is a real advantage no other AI company has. Against that, the flagship models are closed, the data rules are still settling, and the product line changes monthly.
If you are deciding which AI tools your team should standardize on, start with the work, not the model: which workflow costs the most hours, where the data lives, and what your privacy guidelines allow. An AI readiness assessment answers those questions in a few weeks, and a team workshop gets everyone using the chosen tools safely.
One disambiguation, because it causes confusion in search results: Microsoft also has a model called Muse, announced in February 2025. It is an unrelated research model that generates video game visuals and controller actions, and has nothing to do with Meta's products.
Source: Microsoft Research, Introducing Muse (February 2025)
Wondering whether Muse, ChatGPT, Claude, or Gemini fits your team best? Book a free 30-minute call and we will walk through your workflows and data rules, no pitch attached.
Get Started