Jiberish

The Bench

Meta and Muse: models, plans, and our verdict

Meta's Muse agent, Muse Code, Model API, and open-weight models as of November 10, 2026: what's available, what it costs, and why we rate it Trial.

Kyle Del FranciaUpdated 12 min read
In this piece

Meta on the bench

  • TrialMuse: Muse Wonder, Spark, and Ember, open-weight models, the Muse agent, and Muse Code.Strongest at: Low-cost models and coding agent · Open weights · Consumer agents
See every verdict on the bench

Meta now ships its AI under the Muse name, built by Meta Superintelligence Labs. A business has five things to evaluate: a consumer agent called Muse, Meta AI inside Meta's apps, a terminal coding agent called Muse Code, the pay-as-you-go Meta Model API, and two open-weight models. Sources on this page were checked on November 10, 2026. Meta has shipped something new almost every month since Muse Spark launched on April 8, so check the linked pages before you buy.

Where it stands

Muse is on Trial: strong in places, and worth piloting with review built in. That is our judgment from hands-on testing and Meta's published documentation, plans, and terms.

Meta's range is unusually wide for a lab this new to the market: a consumer agent, Meta AI across Meta's apps, a multi-agent coding tool, a pay-as-you-go API, and two open-weight models. Muse Code is inexpensive to try, with plans from $5 to $200 a month, and the Power and Maximum plans include Muse Wonder Full, which Meta calls its most advanced model. Muse Ember, released November 9, is Meta's smallest and cheapest model, at $0.10 per million input tokens and $0.50 per million output tokens on the API, and is included in every Muse Code plan. Muse Glimmer is open-weight under Apache 2.0, and Muse Spark Open is available under the Muse Model License, which gives teams a path to self-hosting. The Muse agent has sensible guardrails on paper: Meta says each user's agent runs in an isolated cloud machine, passwords sit in a credentials store the model never sees, and Muse asks before sending email or making a purchase (Muse privacy help).

Three things keep it from Adopt. The data defaults lean toward training: Muse uses interactions to develop AI models unless you switch that off, and Muse Code subscriptions default to Contributor mode, under which Meta can train on prompts, code, and files, while Meta's Model API terms say sensitive or confidential information must not go to Contributor services. The products are young, with Muse Wonder still in preview. And the Muse agent has no business plan with admin controls yet.

We would move Muse to Adopt when the agent gains a business plan with admin controls, training off by default, and stated retention; when Muse Code offers team plans with Contributor mode off by default; and when Muse Wonder leaves preview. A step back on data terms or availability would move it to Watch.

The lineup today

SurfaceWhat's thereStatus on November 10
Muse agentPersonal agent on iOS, Android, muse.ai, Mac, and in WhatsApp; launched on Muse Spark; voice with expressive avatars and customizable spaces added in NovemberUS rollout from September 8, US and Canada by September 29, Singapore from October 27; voice and avatars US only to start; glasses in beta
Meta AI in Meta's appsAssistant in the Meta AI app, meta.ai, WhatsApp, Instagram, Facebook, Messenger, and Threads; Muse Image for imagesPowered by Muse Spark 1.1 per Meta's July 24 post, with specialized model versions rolling out from November 6; free for everyday use
Muse CodeMulti-agent terminal coding agent, plus an SDK in developer previewOut of beta since August 31
Meta Model APImuse-wonder-1.0, muse-ember-1.0, muse-spark-1.3, muse-spark-1.2, muse-spark-1.1, Contributor variants, muse-image-1.0, muse-voice-transcribe-1.0, sam-3.1Called generally available globally on September 24; Muse Wonder in preview
Open weightsMuse Spark Open (816B-parameter mixture of experts, 42B active, 1M context); Muse Glimmer (30B dense, 128K context by default)Muse Model License; Apache 2.0

Meta's model docs call muse-spark-1.3 the latest version and recommend it for new work. Muse Spark 1.2 and 1.3 each have a cheaper -contributor twin that grants training rights, and muse-spark-1.1 is Standard only and absent from Muse Code. Muse Ember is Standard only, with no Contributor variant. Muse Wonder Full has no API identifier because it is sold only inside Muse Code, neither Wonder model is currently available in South Korea, and Muse Image isn't available in India (Muse Code subscription help). Muse Spark 1.3 is also on Oracle Cloud AI Platform, with Google Cloud in private preview, and Meta says Muse Ember is on Google Cloud and Oracle Cloud. The realtime voice behind the Muse app's new voice mode isn't sold to developers: the model docs say the API doesn't synthesize speech or offer speech-to-speech.

Llama is no longer the headline. Meta's April Muse Spark post calls Llama 4 Maverick "our previous model," and llama.com now redirects to Meta's AI developer site, where the Llama documentation still lists Llama 4 and Llama 3.x with downloads. We found no official statement retiring Llama or naming a successor, so we treat existing Llama deployments as stable but would not start new work on them without a stated roadmap.

Plans and pricing

Prices are in US dollars, as of November 10, 2026, from Meta's own pages. Meta says benefits, pricing, and availability vary by region and account.

Muse agent (Muse subscriptions)

PlanPriceWeekly allowance
Free$0Usage limit that refreshes
Power$20 a month500M Muse tokens
Maximum$100 a month3B Muse tokens

Unused weekly allowance doesn't roll over, and separately purchased tokens don't expire.

Meta AI in Meta's apps. Meta says Meta AI stays free for everyday use. Its Meta One bundles, announced September 15, list Core at $7.99 a month and Premium at $19.99, both adding more Meta AI image and video generation, and the help center says Meta One is in limited testing.

Muse Code (plans)

PlanPriceUsageModels
Everyday$5 a month10 to 50 prompts every 5 hoursMuse Ember, Muse Spark, Muse Wonder, Muse Image
High$15 a month5x EverydaySame as Everyday
Power$50 a month20x EverydayAdds Muse Wonder Full
Maximum$200 a month80x EverydayAdds Muse Wonder Full

Plan contents are from Meta's subscription help, which now lists Muse Ember in every plan. A subscription works only in the Muse Code CLI, other API keys bill pay-as-you-go, and canceled subscriptions aren't refunded unless the law requires it (subscription docs).

Meta Model API, per million tokens (pricing and rate limits)

Model IDInputCached inputOutput
muse-wonder-1.0 (Standard only)$4.00$0.20$20.00
muse-spark-1.3, muse-spark-1.2, muse-spark-1.1$1.25$0.15$4.25
muse-spark-1.3-contributor, muse-spark-1.2-contributor$0.10$0.002$0.20
muse-ember-1.0, up to 200K input tokens (Standard only)$0.10$0.004$0.50
muse-ember-1.0, over 200K input tokens$0.30$0.012$1.50

Muse Image costs $0.01 per image, Muse Voice Transcribe $0.18 per hour of audio, SAM 3.1 $2.50 per 1,000 images, and web search grounding $2.50 per 1,000 queries. Only Muse Ember has a long-context premium: a request over 200K input tokens bills entirely at the higher rate. The pricing page lists no free tier and no batch discount. Meta publishes no fee for the open weights, but Muse Spark Open's 4-bit checkpoint alone is about 537 GB and runs on a four-GPU Blackwell node, per its docs.

Data, privacy, and compliance

The Muse agent trains on your interactions by default. Meta's privacy help says the Data Controls toggle turns that off, including for past interactions, and that Muse doesn't share conversations or virtual machine data with Meta's ad systems. The Muse terms, last updated September 8, say you "will maintain meaningful human oversight over all actions taken" and should not let it send communications or move money without review or pre-authorization.

Meta AI in Meta's apps works differently. Since December 16, 2025, Meta has used Meta AI conversations to personalize content and ads in most regions, excluding sensitive topics such as health and politics from ad targeting. In the EU, Meta trains on adults' public content and Meta AI interactions and offers an objection form; it says it doesn't train on private messages with friends and family, though a March 27, 2026 update says it may train on messages people choose to share with its AI features. Client and customer data should stay out of consumer Meta AI chats.

The Model API's terms, last updated October 9, describe a service intended for commercial use. Meta's data commitments say it does not train on content sent to Standard services, may train on content sent to the discounted Contributor tier, and encrypts traffic with TLS 1.2 or later. Retention is "as needed," with no period stated, per the terms, and Meta began accepting zero-data-retention requests in August. We found no named SOC 2 or ISO certification and no HIPAA terms on the pages we checked. The terms also bar using the service to train other AI models without Meta's written permission and limit use by geography.

In Muse Code, Contributor mode can be turned off in /settings, and an organization can lock it off through managed configuration. Headless runs read each machine's saved setting, so set it explicitly everywhere (subscription docs). Contributor mode isn't available to accounts with zero data retention (subscription help), and on the API it isn't offered for Muse Wonder or Muse Ember, so those models run only under the Standard no-training terms (pricing).

Muse Glimmer uses Apache 2.0. The Muse Model License for Muse Spark Open allows commercial use, fine-tuning, and distillation, but requires "Built with Muse Spark Open" in the product or its documentation, incorporates a usage policy, and requires a separate agreement with Meta if you offer models as a service and your group's gross revenue exceeds $20 million in any 12-month period. Self-hosting either model keeps prompts on infrastructure you control.

What changed recently

  • August 10: Muse Glimmer, a 30B open-weight model (Meta research).
  • August 25: Muse Image on the Model API (blog).
  • August 31: Muse Code left beta with monthly plans and an SDK preview (blog).
  • September 2: Muse Spark 1.3 for agentic and coding work (Meta research).
  • September 8: The Muse agent began rolling out in the US (newsroom).
  • September 15: Meta One subscription bundles (newsroom).
  • September 24: At Connect, Meta called the Model API generally available globally, brought Muse Code to Windows, and put Muse Spark 1.3 on Oracle Cloud (recap).
  • September 29: Muse for Small Business added connectors including Facebook Pages, Meta ad accounts, Shopify, and QuickBooks (newsroom).
  • October 12: Muse Wonder and Muse Wonder Full, plus the Maximum Usage plan for Muse Code (newsroom, blog).
  • October 15: Muse Spark Open (blog).
  • October 27: Muse launched in Singapore (newsroom).
  • November 5: Voice conversations and expressive avatars began rolling out in the Muse app on mobile, web, and Mac, in the US to start (newsroom).
  • November 6: Spaces, custom experiences you design inside the Muse app, available wherever Muse is (newsroom), and specialized versions of Meta's existing models, trained using Muse Wonder, rolling out to Muse and Meta AI (Meta research).
  • November 9: Muse Ember, Meta's smallest model, on the Model API as muse-ember-1.0 and in every Muse Code plan (Meta research, blog).

Where it fits

The Muse agent suits individuals and small teams who already work in WhatsApp and Instagram, especially marketing-heavy businesses, given its connectors to Facebook Pages, Instagram analytics, and ad accounts. Keep training off, approvals on, and client data and payments out of scope.

Muse Code suits developers who want an inexpensive coding agent for prototypes, internal tools, and non-confidential repositories, with Contributor mode switched off. On the API, Muse Spark in Standard mode is priced for high-volume extraction and drafting, and Muse Ember undercuts it for routing, tool calls, and coding subtasks handed off by a larger model. Muse Glimmer suits features that must run on a device or offline, and Muse Spark Open suits organizations that must self-host and can accept its attribution clause and hardware bill.

It is not yet a fit for regulated health or financial data, confidential code under Contributor mode, or customer-facing automation that needs organization-level admin and audit.

What we're watching

  • Muse Wonder itself inside the Muse agent, which Meta says will reach subscribers first, and Muse Wonder leaving preview on the API. The November 6 models are trained using Muse Wonder but are not Muse Wonder.
  • The full Muse Wonder safety report.
  • Muse Confidential VM, which Meta said on September 8 would arrive "later this year" to keep data private even from Meta.
  • A business plan for the Muse agent and team billing for Muse Code.
  • Muse Video, which Meta previewed on July 7 as coming soon.
  • Muse on AI glasses leaving beta, plus the promised Shop Pay and 1Password support.
  • Voice conversations reaching AI glasses and Muse Charm, which Meta says are coming soon, and voice and avatars outside the US.
  • Muse Charm, the pocket device for Muse shown at Connect, where Meta said it would have more to share later this year.
  • Shared spaces, which Meta says it is exploring.
  • An official statement on the Llama roadmap.

We will update this page as those land. Compare Meta with the other labs on the Bench, including Anthropic and OpenAI.

From verdict to deployment

Putting Meta to work is the hard part.

We design and build AI into the websites and applications a business runs on, automate workflows with agents that work across its tools, and roll AI out across whole organizations, with every team trained on its real work. We build on whichever lab fits the job, and we stay through launch and adoption.

Book a free call

Put this into practice

Bring the decision you’re working through.

A free discovery session is a chance to review your goal, the tools or website you have, and the help you need. We’ll identify a sensible next step and discuss scope before any project commitment.

AI Advisory & Training →AI-Embedded Apps & Portals →Book a free call