What is Meta Muse Ember and what does it cost?
Meta released Muse Ember on Nov. 9, 2026, the fourth model in its Muse family and the smallest the company has shipped. Meta Superintelligence Labs built it for “complex reasoning execution and the work agents repeat all day”: tool calls, coding subtasks, routing, extraction and long document parsing.
The list price is $0.10 per million input tokens and $0.50 per million output tokens, with cached input at $0.004. Meta pairs that rate with a one-million-token context window and up to 512,000 output tokens, enough to load a codebase or a document set in a single call.
Meta says the model handles text, images, audio and video natively, so an agent can read screenshots, charts and scanned documents or work from meeting recordings without a separate model. OpenRouter's listing describes the same inputs, including PDF documents.
Developers can reach it through the Meta Model API as muse-ember-1.0, through Muse Code, and on Google Cloud and Oracle Cloud. It supports the Responses API. Prompts above 200,000 tokens cost more: $0.30 per million input and $1.50 per million output, with cached input at $0.012.
Ember follows Muse Spark in April, Glimmer in August and Muse Wonder in October. Meta says Ember is the first model trained using Muse Wonder Full, its largest system, and describes the goal as capability “at a price that scales.”
The family is tiered sharply. Meta's own comparison puts Muse Spark 1.3 at 37 on the same index for a listed $1.25 input and $4.25 output per million tokens, which Artificial Analysis prices at $2.39 per task. Ember gives up one index point but costs about one-seventeenth as much per task.
Why a per-task price matters more than a per-token price
Agents rarely make one model call. They make hundreds, and most of those calls are small: read a file, check a schema, route a ticket, retry a failed step. Cost per task, not cost per token, decides whether that workflow is affordable at scale.
Meta leans on exactly that framing. Its announcement sums up the design as “Muse Wonder plans, Muse Ember handles the volume.” A heavier model sets the rules for a support case; Ember applies them across tens of thousands of cases.
Artificial Analysis, an independent evaluator, computes cost per task by running a fixed set of nine evaluations and weighting the input, cache and output prices each model actually consumed. On that measure, Ember costs about $0.14 per Intelligence Index task. Claude Haiku 5.5 costs $0.41, and GPT-6 Luna costs $0.15.
The striking part is that the sticker prices are nearly identical. Artificial Analysis lists Haiku 5.5 and Luna at the same $0.10 input and $0.50 output rate per million tokens as Ember. The threefold gap on Haiku comes from token use, not from the rate card.
The same evaluator says Ember generated about 115 million tokens while working through the index, against roughly 298 million for Haiku 5.5. Haiku's profile is rated “very verbose,” while Ember's is only “somewhat verbose.” Cheaper thinking, in effect, is cheaper output.
Speed compounds the difference. Artificial Analysis measured Ember at 440 output tokens per second and 3.2 minutes per index task. Haiku 5.5 ran at 254 tokens per second over 14.8 minutes, and Luna at 102 tokens per second over 12.3 minutes.
Cached input deserves as much attention as the headline rate. Ember's cached tokens cost $0.004 per million, a 96% discount on the list price. Agent loops resend the same system prompt, tool schema and retrieved context on every step, so much of the input can bill at the cached rate. Artificial Analysis lists a 90% cache discount for Haiku 5.5, which narrows the gap without closing it.
How does Muse Ember compare with Claude Haiku 5.5 and GPT-6 Luna?
All three models were run at maximum reasoning effort, which favors accuracy over thrift. The headline is a tie: Meta's model and Anthropic's both score 36 on the Artificial Analysis Intelligence Index, while OpenAI's GPT-6 Luna scores 28.
Scores below are listed as Muse Ember, Claude Haiku 5.5 and GPT-6 Luna, all at maximum reasoning effort, using the figures Meta published from Artificial Analysis and its own suite.
- •Artificial Analysis Intelligence Index: 36, 36, 28.
- •AutomationBench-AA, end-to-end business workflows: 37.8, 13.9, 20.2.
- •Agents' Last Exam, end-to-end agent tasks: 61.6, 57.0, 57.1.
- •Terminal-Bench 4.0, agentic terminal coding: 40.9, 39.2, 16.4.
- •Coding Agent Index-AA, coding in real agent harnesses: 46, 37, 31.
- •FrontierCode 2.0, engineering tasks judged on mergeability: 27.2, 34.9, 25.2.
- •GDP.pdf, professional work over long documents: 22.8, 20.8, 22.8.
- •OSWorld 2.1, computer use with partial credit: 69.5, 68.9, 50.5.
- •PerceptionBench with tools, fine-grained visual perception: 80.6, 78.6, 62.8.
Where does that benchmark result actually leave Muse Ember?
The pattern is uneven rather than dominant. Ember leads on business-workflow automation, terminal coding and perception. Haiku 5.5 still wins the benchmark that judges real engineering tasks on whether a patch could be merged, and it matches Luna on long-document work.
Across the nine evaluations Meta published, Ember led on seven, trailed Haiku on FrontierCode 2.0 and tied Luna on GDP.pdf. SiliconANGLE counted seven wins from the same set and singled out the roughly 17-point AutomationBench margin as the model's clearest advantage.
Read the footnotes before treating any of these rows as a ranking. Terminal-Bench 4.0 compares each model in a different harness: mini-swe-agent for Ember, Claude Code for Haiku 5.5 and Codex for Luna. Harness choice can move agentic scores by a wide margin, so that row is not a clean like-for-like comparison.
The internal numbers deserve the same caution. The Personal Superintelligence Index, where Ember scores 46.8 against 36.1 for Haiku, is Meta's own composite of eight evaluations. It is a useful signal about the company's priorities, not an outside verdict.
Is Muse Ember open source?
No. Muse Ember is a hosted model. Meta's launch post lists API and cloud availability but no downloadable weights, and Artificial Analysis classifies it as proprietary.
That breaks with part of Meta's recent pattern. TechCrunch notes that Glimmer, the August release, was an open-source twin of Muse Spark. Ember gets no such release at launch, and Meta has not said whether one is planned.
The closed approach puts Ember in direct competition with open-weight budget models, including DeepSeek V4.1 Flash and GLM 5.3 Flash, which TechCrunch identifies as the closest price comparisons. Those models can be self-hosted, while Ember cannot.
For buyers, the difference is practical rather than ideological. Self-hosting trades a licence fee for hardware, latency and staff. A hosted model trades control for someone else's capacity, rate limits and data terms. Ember currently sits on only the second side of that line.
Meta now sells through Google Cloud and Oracle Cloud as well as its own API. That makes Ember a product for competitors' marketplaces as much as a feature of Meta's own platform.
Who is affected by a cheaper subagent model?
The immediate beneficiaries are teams running high-volume agents: support automation, contract review, claims triage, code review and data extraction. Those workloads spend most of their budget on repetitive calls that need competence more than brilliance.
Platform vendors are already positioning around that use. Meta published test notes from Box, Cline, Oracle, Cloudflare and You.com. Cline's founder, Saoud Rizwan, said Ember is “purpose-built for, and excels at, subagent work like coding and document analysis.”
Cloudflare's chief technology officer, Dane Knecht, said the model “ran up to 30% faster at the tail than the comparable model we tested.” You.com reported a p95 latency of 2.2 seconds and 36% to 48% fewer output tokens than leading open-weight models.
Box's vice president of AI products, Yashodha Bhavnani, called it “a fast, efficient foundation for agents doing real work on enterprise content.” These are vendor-selected testimonials from early testing, not independent procurement reviews.
The competitive effect lands on Anthropic and OpenAI. If buyers start comparing models by measured cost per finished task, a tie on intelligence at a third of the spend changes routing defaults inside agent frameworks. It also pushes smaller models toward token efficiency, not just lower posted rates.
Procurement teams have a second reason to care. A model that matches a frontier-adjacent score at a lower per-task cost is easier to defend in a budget review than an upgrade that lifts benchmark points. Expect more requests for task-level cost evidence, and fewer rate-card comparisons.
For Meta, Ember completes a ladder: Wonder for hard problems, Spark for coding and agents, Ember for volume. Meta says the cost is low enough to run “in every product, for everyone.”
What does Muse Ember signal about Meta's model business?
Meta spent years positioning open weights as its contribution to the wider market. Ember, unlike August's Glimmer, is sold rather than given away. The company is now a model vendor competing on price against the labs it once treated as peers in an open ecosystem.
Distribution is its edge. Meta's consumer agent, Muse Agent, arrived in September and has picked up features such as Spaces and video chat this month, according to TechCrunch. A model priced at roughly 14 cents per task is cheap enough to sit inside consumer products without an obvious meter running.
The cloud partnerships cut both ways. Selling Ember on Google Cloud and Oracle Cloud buys distribution beyond Meta's own API. It also puts Meta's model beside Google's and Oracle's own offerings, where buyers can compare them on the same invoice.
The low end of the market is where the fight is sharpest. Haiku 5.5 and GPT-6 Luna already sit at the same list price as Ember, and open-weight rivals such as DeepSeek V4.1 Flash and GLM 5.3 Flash undercut on licence cost. Meta's argument is that its token efficiency and speed make the total bill smaller.
What are the limitations and open questions?
Most of the evidence is vendor-published or vendor-commissioned. Meta chose the benchmarks, ran the internal ones and supplied the cost chart. Artificial Analysis provides an independent index, but it is one composite, not a guarantee for any particular workload.
A tie is not a win. Ember matches Haiku 5.5 at 36 on that index; it does not beat it. The FrontierCode 2.0 result, where Haiku leads by nearly eight points, is the clearest sign that Ember is a specialist rather than a general upgrade.
The rankings add a little modesty too. Within Artificial Analysis's under-$0.15 price class, Haiku 5.5 is ranked first of 33 models on intelligence and Ember second. On cost, Ember ranks eleventh. It is excellent value, but it is not the cheapest or the smartest model in its bracket.
Cost per task also depends on assumptions. Artificial Analysis weights evaluations and token mixes in a fixed way, while real prompts skew long or short, cached or fresh. Ember's price triples for prompts above 200,000 tokens, which is easy to hit with large codebases.
The multimodal list needs checking per use case. Meta and OpenRouter list audio and video inputs, while Artificial Analysis currently lists text and image only. Teams that need audio should verify support in their own account before designing around it.
Operational details are thin. The launch post gives no rate limits, uptime commitment or regional data-residency options. For regulated buyers, those gaps matter more than a benchmark point.
Safety testing was Meta's own. The company says Ember stayed below high-risk thresholds under its Superintelligence Scaling Framework, including tests for hidden instructions in pages, documents and tool outputs. That is a useful disclosure, not an external audit.
Verbosity is a two-sided result as well. Ember generated far fewer tokens than Haiku, but Artificial Analysis still rates it somewhat verbose against the median model.
What should developers watch next?
The next signal is independent reproduction. Meta's biggest margin sits on AutomationBench-AA, a business-workflow benchmark, and third-party runs on private tasks will show whether that lead survives outside the launch chart.
Watch pricing at Anthropic and OpenAI. Both competitors already list $0.10 input and $0.50 output per million tokens, so a response would more likely come as faster variants or leaner token use than as a deeper rate cut.
Watch Meta's roadmap too. The company says it has “even more efficient models in training,” and it has released open weights for a Muse sibling before. Either move would change the calculus for teams that want to self-host.
For now, the practical test is narrow. Run your own agent traces through Ember and your current model, measure tokens and wall-clock time per completed task, then compare invoices. That number, not the leaderboard, decides the routing table.
