Meta Muse Ember: The $0.10 Model for High-Volume AI Agents

Meta released Muse Ember on November 9, 2026, and the headline number is hard to miss: $0.10 per million input tokens. It is the smallest and cheapest model in Meta's Muse family, built for the repetitive work AI agents do all day: tool calls, coding subtasks, routing, extraction, and long document parsing. The model is live now on Muse Code and the Meta Model API as muse-ember-1.0, with a 1M-token context window and up to 512K output tokens.

There is a catch in the headline, and it matters more than the price itself. Ember's $0.10 input rate and $0.50 output rate are the exact list prices of Claude Haiku 5.5 and GPT-6 Luna, verified on each vendor's own pricing page. Meta's saving is measured per task, not per token: the company puts Ember at about $0.14 per Artificial Analysis task and says it matches Haiku's best benchmark score at about a third of Haiku's cost per task. That is a real advantage if your agents run thousands of small steps a day, and a much smaller one if they do not. This article works through the price, the benchmarks, and the places where Meta's own numbers favor the competition.

What Meta actually launched

Ember is the fourth model under Meta's Muse brand. The line started with Muse Spark, a mid-weight model released in April 2026, followed by Glimmer, its open-weight twin, in August, and the frontier model Muse Wonder in October. The flagship consumer product, the Muse Agent, arrived in September. Ember rounds out the portfolio as the small, cheap workhorse: Wonder for the hardest problems, Spark for coding and agentic workflows, and Ember for repeatable coding and agentic tasks at scale.

Meta describes Ember as "designed for complex reasoning execution and the work agents repeat all day." An agent's bill is dominated by many small, structured steps, so a model built for this job needs to be cheap per step, fast to call, and reliable at tool use. Its 1M-token context window takes in a full codebase or document set in one call, and the 512K-token output ceiling means long generations are not split across calls.

The model is multimodal by design. It natively understands text, images, audio, and video, so agents can read screenshots, charts, and scanned documents, or work from voice memos, meeting recordings, and video without a separate model. Box, an early partner, tested Ember on its Box AI Complex Work Eval. Yashodha Bhavnani, Box's VP of AI Products, said the model "can score a contract against a company's negotiation playbook clause by clause and land on the right overall risk rating."

Ember also carries an unusual piece of lineage: Meta says it is the first model trained using Muse Wonder Full, the company's most capable model, which sits behind Muse Code's $50 and $200 subscription tiers and is not sold per token at all. When Wonder launched, Meta said it would use it to develop specialized models for everyone. Ember is the first result, and Meta says more efficient models are already in training.

What it costs

Ember is available on the Meta Model API at three rates, all in US dollars per million tokens:

The arithmetic is straightforward. A request with 1 million input tokens and 250,000 output tokens costs $0.10 plus $0.125, or $0.225. The same request on Muse Spark 1.3, at $1.25 input and $4.25 output, costs $2.31, about ten times as much. On Muse Wonder, at $4 input and $20 output, it costs $9.00, forty times Ember's price. Those are Meta's own published rates, and the gap is the whole point of the model: Ember exists so the cheap steps in an agent workflow stop eating the budget.

For readers outside the US, the dollar figures convert to roughly £0.07 and CA$0.14 per million input tokens, and roughly £0.37 and CA$0.71 per million output tokens, at November 10, 2026 market rates (1 USD = 0.7454 GBP, 1.4103 CAD). Meta bills in US dollars, so treat these as approximations. The often-quoted $0.14 per task works out to about £0.10 or CA$0.20.

The other way to reach Ember is Muse Code, Meta's terminal coding agent, where the model is available from today. Muse Code subscriptions start at $5 a month, with higher tiers at $15, $50, and $200, and Meta lists no free tier, according to independent trackers that read Meta's product pages. The API route needs no subscription: you pay per token, with no monthly fee. Subscriptions default to Contributor mode, which trades more usage for letting Meta use your prompts, code, and files to train future models. Switch it off before your first prompt on private code.

The same sticker price as its rivals

Here is the comparison Meta's launch announcement invites, with every rate taken from the vendors' own pricing pages on November 9, 2026:

* Haiku 5.5's $0.10 input and $0.50 output rates apply only to prompts up to 100,000 tokens. Above that line, Anthropic charges $0.50 input and $2.50 output, five times as much, for the whole request. Luna has its own cliff: prompts over 272,000 input tokens are billed at $0.20 input and $0.75 output for the full request. Ember's announcement lists flat rates with no long-prompt surcharge, which is a genuine edge for long-document work, though Meta has not said the rates can never change.

Two details in that table do the real work. First, the three small models share the same $0.10 and $0.50 list prices, so switching from Haiku or Luna to Ember on sticker price alone means switching to the same price. The saving comes from token efficiency: Ember uses fewer tokens to finish the same benchmark task, which is why Meta's cost-per-task figure is about a third of Haiku's. Second, Ember's cached input is the cheapest of the three at $0.004 per million tokens, against $0.01 for both rivals. On 10 million cached tokens, Ember costs $0.04 where Haiku and Luna cost $0.10, a gap that adds up fast for agents resending the same long prompt on every step.

Ember also sits far below its own siblings: Spark 1.3 costs $2.39 per Artificial Analysis task against Ember's $0.14, about seventeen times as much, while scoring only one point higher on the same index, 37 against 36. Wonder costs $4.52 per task. Ember is not the smartest Muse model. It is the one built to be run constantly.

What the benchmarks say, including the one Ember loses

Meta's announcement leans on Artificial Analysis benchmarks dated November 9, 2026, measured at each model's maximum reasoning effort. The headline figures are strong, and the company says Ember "delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting" and that "no model scores higher for less per task" across the full leaderboard. Both claims are Meta's framing of its own chart. The scores come from Artificial Analysis, not from an independent reproduction, and the tests below are Meta-selected.

The wins are real and concentrated where Meta aimed the model. On AutomationBench, which runs end-to-end business workflows, Ember scores 37.8 against Haiku's 13.9, nearly three times as much. On the Coding Agent Index it leads 46 to 37, and on Terminal-Bench 4.0 it edges Haiku 40.9 to 39.2. On OSWorld computer use it leads on both scorings, 69.5 and 33.8 against Haiku's 68.9 and 27.8.

But the table also contains the result Meta would rather you skim past. On FrontierCode 2.0, which asks whether a human maintainer would merge an agent's code, Ember scores 27.2%, behind Haiku's 34.9% and only just ahead of Luna's 25.2%. That is the benchmark closest to real engineering work, and it is the one where Ember trails. Ember is better at running workflows and following instructions than at producing code a reviewer would accept unchanged, and if your agent's output goes straight to a human reviewer, that gap matters more than the AutomationBench lead.

The Artificial Analysis Index itself is a tie at 36 apiece, with Luna at 28. Meta's cost claim rests on that tie: the same rounded score at about a third of Haiku's cost per task. But a cost-per-task figure depends on how many tokens each model burns, which varies with the task and the prompt. The only way to know your own number is to run your own workload on both.

Where Ember fits: Wonder plans, Ember handles the volume

Meta's positioning line for the model is blunt: "Muse Wonder plans, Muse Ember handles the volume." The pattern is a two-model harness. A heavy model establishes the rules for handling an issue, then a lighter model applies those rules across tens of thousands of cases. In a customer-support agent, for example, Wonder would work out how a particular problem should be handled, and Ember would apply that handling to every incoming ticket. The same shape fits contract review, candidate screening, document triage, and any workflow where the same judgment repeats at scale.

That is also why the 512K-token output ceiling matters. Luna stops at 128,000 tokens, so Ember's ceiling is four times higher: one call can produce a long report or a large code change without being split into chunks. Whether quality holds up across half a million output tokens is untested, and long outputs are where small models usually degrade first. The multimodal reach saves calls the same way: an agent that reads screenshots and voice memos without a second model saves money on every step.

The catches worth knowing before you switch

First, there is no open-weight version. Meta announced no weights release for Ember, and the model is available only through Muse Code and the Meta Model API. Glimmer is the open twin in the Muse family, and it belongs to Spark, not Ember. If you need to run a model on your own hardware, Ember is not the answer, whatever its price.

Second, there is no consumer route yet. The announcement covers Muse Code and the Meta Model API only; the consumer Muse app is not mentioned at all. Anyone waiting to use the cheapest Muse model inside the app they already have is waiting on a timeline Meta has not published.

Third, the safety claims are Meta's own. Before release, the company evaluated Ember under its Meta Superintelligence Scaling Framework, covering chemical and biological, cybersecurity, and loss-of-control risks, and says the model stayed below the framework's high-risk thresholds in the categories assessed. It also says Ember received the same core safety training as its larger siblings, resists instructions hidden in web pages, documents, or tool outputs that try to redirect it, and seeks human input before high-impact actions. No independent audit has been published, and agent models face prompt-injection attempts constantly in real deployments. Treat the safety section as a promise to verify, not a result.

Who should switch, and who should wait

Switch, or at least test, if your agents run high-volume repetitive steps. Tool-calling loops, routing, extraction, and long-document parsing are the workloads Ember was built for, and the AutomationBench and Coding Agent Index leads suggest it will hold up there. Test it first if your prompts are long: Haiku's price jumps fivefold above 100,000 tokens and Luna's doubles above 272,000, while Ember's rates stay flat across its 1M-token window. If your agents resend the same context on every step, Ember's $0.004 cached-input rate is the cheapest of the three by a clear margin.

Wait if merge-quality code is the bottleneck. FrontierCode is the benchmark closest to shipping real code, and Haiku leads it by nearly eight points. Wait too if you need open weights or the consumer Muse app: neither is available, and neither has a date. And do not switch on the headline alone. At $0.10 per million input tokens, Ember costs exactly what Haiku 5.5 and GPT-6 Luna cost. The saving is in how many tokens each model needs to finish your task. Run your own workload on two models for a week before moving anything that matters.

Frequently asked questions

How much does Muse Ember cost?

$0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens on the Meta Model API. Meta puts the cost at about $0.14 per Artificial Analysis task. API access needs no subscription; Muse Code starts at $5 a month with no free tier.

Is Muse Ember cheaper than Claude Haiku 5.5?

Per token, no: both charge $0.10 per million input tokens and $0.50 per million output tokens for short prompts. Per task, Meta says Ember matches Haiku's best benchmark score at about a third of Haiku's cost, because it uses fewer tokens per task. Haiku's rates also jump fivefold on prompts over 100,000 tokens, while Ember's stay flat.

Is Muse Ember open source?

No. Meta announced no weights release, and Ember is available only through Muse Code and the Meta Model API. The open-weight model in the family is Glimmer, the twin of Muse Spark.

What is Muse Ember's context window?

1 million tokens of input, with up to 512,000 output tokens. That is roughly the same input window as GPT-6 Luna's 1.05 million tokens, but Ember allows four times Luna's 128,000-token output ceiling.

Can I use Muse Ember in the Muse app?

Not yet. Meta's announcement covers Muse Code and the Meta Model API only, with no timeline for a consumer app release.

What does Muse Ember do best?

The repetitive steps inside agent workflows: tool calls, coding subtasks, routing, extraction, and long document parsing. Its strongest results are on business workflow automation and coding-agent harnesses. It is weaker on FrontierCode, where a human decides whether to merge its code, and Haiku still leads there.

Meta's cheapest model is also its most honest test of a simple idea: that the expensive part of AI agents is not the clever thinking, but the thousands of small steps around it. Ember prices those steps at a tenth of what the mid-weight models charge, matches its rivals' benchmark scores while doing it, and loses to them exactly where a human still has to check the work. If your agents do volume, test it this week. If they do judgment, keep Haiku or Wonder on the job and let Ember handle the rest.

Sources

• Introducing Our Smallest, Most Efficient Model: Muse Ember — Meta (official announcement, November 9, 2026)
• Meta releases Muse Ember, a new lightweight model for agent execution — TechCrunch, November 9, 2026
• Claude platform pricing — Anthropic (Haiku 5.5: $0.10/$0.50 up to 100K tokens, $0.50/$2.50 above)
• OpenAI API pricing — OpenAI (GPT-6 Luna: $0.10/$0.50, $0.01 cached)
• GPT-6 Luna model page — OpenAI (1.05M context, 128K max output, 272K long-prompt tier)
• Muse Wonder: benchmarks, pricing and training on your code — Le Recul (Wonder $4/$20, $4.52 per task, 1M context, Contributor mode terms)
• Muse Code vs Claude Code vs Codex in 2026 — Enjoy.dev (Muse Code plans $5/$15/$50/$200, no free tier, Contributor mode; checked October 31, 2026)
• USD to GBP market rate, November 10, 2026 — RateMorph (1 USD = 0.7454 GBP)
• USD to CAD market rate, November 10, 2026 — RateMorph (1 USD = 1.4103 CAD)
• Meta's Muse Spark Open: 816B AI Model You Can Download Free — AI Core Updates
• Claude Haiku 5.5: 90% Cheaper Than Haiku 4.5 — With Two Catches — AI Core Updates
• DeepSeek V4.1 Flash 1017: The $0.15 Coding Model and Its New Search API — AI Core Updates