Does Muse Ember change Meta Muse for everyday tasks?

What is Muse Ember, and what did Meta actually release?

Muse Ember is Meta's smallest and cheapest Muse model. Meta Superintelligence Labs released it on November 9, 2026 to handle the repetitive work inside AI agents. It isn't a new chatbot, a new app or a new button in WhatsApp. Meta says it's built for the jobs agents repeat all day: calling tools in sequence, routing requests, pulling details out of documents, parsing long files and handling small coding subtasks. Developers can use it now through Muse Code and Meta's model API. It's also the fourth model to carry the Muse name. Spark came in April, its open-source twin Glimmer followed in August, and the frontier-sized Wonder arrived in October. If you were hoping for something new to tap in your chats this week, there isn't anything.

The interesting part is how Meta expects it to be used. Ember isn't supposed to do all the thinking. It's meant to sit next to a heavier model like Muse Wonder inside an agent. Meta's own example is a customer-support agent. The big model works out the rules for handling a tricky issue, and Ember then applies those rules across tens of thousands of cases without burning through the budget. Meta also pitches it for fan-out, where one agent spins up several smaller helpers that each tackle part of a job in parallel. Picture a manager who writes the checklist and then hands it to a fast, inexpensive crew. Ember is the crew: quick, cheap and good at following clear instructions.

The specs back up that pitch. Ember takes text, images, audio and video as input. It has a one-million-token context window, which is enough to hold hundreds of thousands of words in a single task, and it can produce up to 512,000 tokens of output. The headline is the price: 10 cents per million input tokens and 50 cents per million output tokens. A token is roughly a chunk of a word, so those numbers sit far below what frontier models charge. For developers, that's the whole story. When every step of an agent costs a fraction of a cent, you can afford to let it check, retry and double-check instead of rationing each call like it's the last biscuit in the tin.

Does Muse Ember change how Meta Muse works on WhatsApp?

Not directly, at least not yet. Muse Ember is a model for developers building agents, and Meta hasn't said it now powers the Muse assistant people chat with in WhatsApp. That consumer product is a different thing. Muse is Meta's personal agent. It launched in the US in September 2026 for adults 18 and over, and it's built to take actions like booking, buying and filling in forms for you. Reports put WhatsApp access in the US and Canada, with the wider Muse service also available in a few other countries. You set it up through a Muse account first and then switch on WhatsApp messaging. Coverage of the launch said it runs on Muse Spark 1.3, not Ember.

Could that change? Sure. A model this cheap is exactly what you'd use for the small, repetitive steps inside a consumer agent: reading a confirmation email, pulling a date out of a PDF, checking a form field. If Meta eventually routes those steps through Ember, Muse might get faster, or the free tier might stretch further. That's a reasonable guess, but it's still a guess, and Meta hasn't announced anything like it. Some things don't change either way. Muse is still limited to certain countries, still needs its own account, and is still separate from the Meta AI assistant already built into WhatsApp. If you're outside the supported countries, here's what to use when Muse isn't available where you live.

Here's the bit that gets lost in launch-day coverage: swapping one model for another inside a product is invisible to you. You won't get a notification saying "now running on Ember". What you actually feel are product decisions. Where does the assistant live? Does it remember what you told it last Tuesday? Does it fire the reminder on time, and does it check with you before buying something? Those are design choices, and a new model doesn't make them for anyone. If you want to see how Muse holds up on everyday chores, we've put Muse Spaces through real reminders and family lists. That's a far better predictor of your Tuesday than any benchmark.

Muse Ember vs Haiku: does the cheaper model actually win?

On overall scores, Muse Ember ties Claude Haiku 5.5 at roughly a third of the cost per task. A widely followed independent intelligence index gives both models a score of 36. Ember's cost works out to about 14 cents per task in that testing. For comparison, Meta's larger Muse Spark 1.3 scored 37 on the same index but cost $2.39 per task. That's nearly identical intelligence for a small fraction of the price. It's also why developers are paying attention: when you run thousands of agent steps a day, the gap between 14 cents and over two dollars stops being a rounding error and starts being a real budget line.

The detail is messier than the headline, though. On Terminal-Bench 4.0, which tests whether a model can get real work done in a command-line environment, Ember scored 40.9% against Haiku's 32.8%. That's a clear win. On FrontierCode 2.0, a tougher coding test, the order flips: Haiku scored 34.9% and Ember 27.2%. Early reviewers noted the same pattern, saying Ember falls short on actual programming even though it's strong at agent-style tool use. Meta also says Ember beats Haiku 5.5 and GPT-6 Luna on several other benchmarks, including OSWorld 2.1, which measures how well a model operates a computer. So it's a strong doer and a weaker coder, which is roughly what you'd want from a model built to carry out instructions.

Now the honest take: for most people reading this, none of it matters much. Price per million tokens is a developer's line item. You pay for an assistant, not for tokens, and terminal benchmarks don't tell you whether "Remind me to renew the van insurance on March 3" will actually fire on March 3. There's even a mild warning hidden in the good news. As agents get cheaper to run, companies will let them take more steps on their own, and every extra unsupervised step is another chance to do something you didn't ask for. Cheaper models make an agent's manners more important, not less. Judge an assistant by how it behaves in your chat, not by its spec sheet.

What should you check before trusting an AI agent with daily admin?

Before you hand any AI agent your daily admin, check three things: does it work where you already chat, does it ask before acting, and does it report back. Start with the first one. An assistant you have to remember to open is an assistant you'll stop using by week three. Most small business owners already run their day out of WhatsApp. That's where clients confirm appointments, suppliers send invoices and the team grumbles about the rota. An agent that lives in a separate app or browser tab asks you to build a new habit, and habits are expensive. One that sits in your chat list like any other contact asks for nothing. We used a similar lens when we looked at how well the Wabi agent handles everyday admin.

The second check matters more every month: does the agent ask before it acts? Reminding you of something is low-risk. Sending a message to your landlord, booking a table or buying a replacement part isn't. A well-behaved agent shows you what it's about to do and waits for a "yes". Muse, to its credit, launched with approval gates for purchases. Apply the same test to anything that talks to other people for you. Can you see the exact message before it goes out? Can you say "make it friendlier" or "shorter, no emoji" and have that stick? Does the agent say clearly who it is, or does it pretend to be you? If an assistant can't answer those three questions cleanly, keep it away from your clients.

The third check is the one people forget until something goes quiet: does it report back? Delegation only works if the loop closes. You ask the agent to find out when the plumber can come, and later you need to know if it got an answer, got a no, or got silence. A good assistant tells you without being chased, and it also answers when you ask, like "Did Dana answer yet?" or "What happened with the plumber?". Watch for silent failures too. A reminder that never fires and never tells you it failed is worse than no reminder at all, because you stopped tracking the thing yourself. Test any agent on a few low-stakes tasks for a week before you trust it with a deadline that costs money.

What can a WhatsApp assistant handle for your admin today?

A WhatsApp-native assistant like Agent Sonic already handles reminders, to-dos, notes, documents and simple errands. You just message it like any other contact, and there's nothing to install. You text "Remind me to call the accountant every Monday at 9" and the reminder turns up in the same chat every Monday. You send "Add buy printer ink to my list" from the car and it's there when you need it. Sonic also remembers what you tell it, so "My VAT number is in the notes, what is it again?" actually works. Forward a lease or contract PDF and ask "What's the notice period?". Send a voice note when typing isn't an option. It can also add events to Google Calendar or update a Google Sheet straight from the chat.

For errands that involve other people, Sonic has outreach tasks, available on the Pro plan and currently in Beta, so they may be paused from time to time. You say something like "Ask Dana when the invoice will be ready and let me know". Before anything is sent, Sonic shows you the exact first message. You reply yes, no, or a change like "more formal". It writes from its own WhatsApp number, introduces itself as your assistant, follows up at most once if the answer is vague, and reports back when the thread ends. The limits are deliberate: one task at a time, up to 3 people a day, first messages only between 08:00 and 21:00 in the recipient's time zone, and anyone who says stop is never contacted again. Here's how outreach tasks work step by step.

Sonic also works inside WhatsApp groups, which is handy if you coordinate a team, a flat share or a five-a-side squad. Add it to a group to get a recap of what you missed, run a quick poll, keep a shared list, split expenses or set automatic reminders, like "Every Friday at 4, remind everyone to submit timesheets". So if your main question about Muse Ember was "which assistant should run my errands?", the honest answer is that the model is the least important part. Pick the one that lives in your chats, asks before acting and closes the loop. If that sounds like what you need, get access to Sonic: leave your details and you'll get Sonic's number and a quick onboarding. Plan details are on the pricing page.

Frequently asked questions