Takeaway: Meta Superintelligence Labs released Muse Ember on November 9, 2026, calling it the lab’s smallest and most efficient model for the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long-document parsing. On Artificial Analysis’s Intelligence Index, Meta said Ember at maximum reasoning effort matches Anthropic’s Claude Haiku 5.5 at about a third of the cost per task — 35.8 versus 36.2, at an estimated $0.136 versus $0.411 — and beats every published GPT-6 Luna setting. Both rounded Index scores are 36. The model is available now on the Meta Model API and in Muse Code as muse-ember-1.0, at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 200,000 tokens. Longer prompts cost more. Several of the standout scores in Meta’s post are the company’s own evaluations, and Haiku still leads on FrontierCode 2.0, which grades whether changes look ready to merge.
This is a developer-facing model launch, not a consumer Muse app update. It is distinct from last week’s Spaces customization and Active Memory models inside the Muse app, and from the U.S. Realtime Voice and Avatar rollout. TechCrunch’s Russell Brandom placed Ember as the fourth model under the Muse brand, after the mid-weight Spark (April), Spark’s open-weight twin Glimmer (August), and the Wonder frontier model (October).
What Meta says Ember is for
Meta’s framing is not “a tiny chatbot.” The company wrote that Ember is built for complex reasoning execution — the many small, structured steps that dominate an agent’s token bill. In a support workflow, that might mean looking up records, reading documents, and calling tools before a heavier model, or a human, decides what to do. Meta’s shorthand, repeated in the launch post, is: “Muse Wonder plans, Muse Ember handles the volume.”
VentureBeat’s Carl Franzen, writing later the same day, made the same point with a ticket-desk example: a single case can send an agent through a chain of lookups and tool calls before it answers. Ember is meant to take those steps, and jobs such as running tests or making small code fixes, at a price that can be repeated thousands of times. Meta also said Ember natively understands text, images, audio, and video, so an agent can read screenshots, charts, and scanned documents, or work from voice memos, meeting recordings, and video, without a separate model. VentureBeat added that the API accepts PDFs and returns text.
Price, the 200,000-token cutoff, and why list rates are not the whole bill
The headline rate in Meta’s post is $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens, with a 1 million-token context window and up to 512,000 output tokens. A footnote matters as much as the headline. For prompts longer than 200,000 tokens, Muse Ember costs $0.30 per million input tokens, $0.012 per million cached input tokens, and $1.50 per million output tokens. VentureBeat reported that the higher rates apply to the entire request, including its output, and that cached input and conversation history carried forward through previous_response_id count toward the cutoff. In its example, a request with 200,000 uncached input tokens and 10,000 billed output tokens costs 2.5 cents; one more input token takes that same request to about 7.5 cents.
That cutoff is the practical difference versus close rivals that start at the same $0.10 / $0.50 list price. VentureBeat noted that Anthropic’s standard Haiku 5.5 rates rise to $0.50 input and $2.50 output above 100,000 input tokens, while OpenAI’s GPT-6 Luna long-context rates are $0.20 input, $0.02 cached input, and $0.75 output. A workload with large prompts and long answers can therefore favor Luna on token rates even when Ember looks cheaper on Artificial Analysis’s benchmark suite. VentureBeat, citing Artificial Analysis, also said Ember used roughly a third as many output tokens as Haiku at maximum effort, and that cheaper cache reads help the cost-per-task figure. Those are reasons a production bill can diverge from a leaderboard, not a promise that every job will land at $0.14.
Against Meta’s own mid-weight model, the cost gap is larger still. Meta’s comparison table puts Muse Ember at an Intelligence Index of 36 and about $0.14 per Index task, versus Muse Spark 1.3 at 37 and $2.39, on list prices of $0.10 / $0.50 versus $1.25 / $4.25 per million tokens. Ember is slightly behind Spark on that Index and far cheaper to run. Meta’s close is that frontier capability has to ship at a price that scales: “We have even more efficient models in training and look forward to sharing what’s next.”
The Intelligence Index comparison, with Meta’s own numbers
Meta published Artificial Analysis Intelligence Index scores against estimated cost per task at each reasoning-effort setting, dated November 9, 2026, and using list pricing. The points below are from that chart. Ember’s two published settings are xhigh and max. Haiku and Luna show more effort steps. “About a third of the cost” is Meta’s summary of Ember’s best score versus Haiku’s best score; the unrounded figures are 35.8 at $0.136 versus 36.2 at $0.411.
| Model and setting | Intelligence Index score | Estimated cost per task |
|---|---|---|
| Muse Ember, xhigh | 33.9 | $0.116 |
| Muse Ember, max | 35.8 (rounded 36) | $0.136 |
| Claude Haiku 5.5, lowest published setting in the chart | 24.6 | $0.092 |
| Claude Haiku 5.5, mid settings | 29.6 and 34.3 | $0.159 and $0.248 |
| Claude Haiku 5.5, max | 36.2 (rounded 36) | $0.411 |
| GPT-6 Luna, lowest published setting | 21.4 | $0.073 |
| GPT-6 Luna, mid setting | 23.3 | $0.104 |
| GPT-6 Luna, max | 27.8 | $0.152 |
Meta also said Ember “sits on the efficiency frontier of the full Artificial Analysis leaderboard: no model scores higher for less per task.” That is the company’s reading of the board on launch day, not an independent audit we re-ran. Artificial Analysis’s own model pages list Muse Ember (Max) at 36 on the Intelligence Index, about $0.14 per Index task, a 1 million-token context window, and the same $0.10 / $0.50 list price. We have not independently reproduced the Index.
Where Ember leads — and where it does not
Meta’s evaluation table mixes Artificial Analysis results with internal scores. The figures below are as published in the launch post, at maximum reasoning effort. Several rows are Meta’s own tests and should not be read as third-party benchmarks. Terminal-Bench 4.0 used each model’s best published harness: mini-swe-agent for Ember, Claude Code for Haiku, and Codex for Luna, so those coding scores are not a same-harness comparison.
| Evaluation | Muse Ember (max) | Claude Haiku 5.5 (max) | GPT-6 Luna (max) | Source, as labeled by Meta |
|---|---|---|---|---|
| AutomationBench-AA, end-to-end business workflows (strict pass@1) | 37.8% | 13.9% | 20.2% | Artificial Analysis |
| Agents’ Last Exam, Linux subset | 61.6% | 57.0% | 57.1% | Meta internal |
| Personal Superintelligence Index | 46.8% | 36.1% | 37.0% | Meta internal composite |
| Terminal-Bench 4.0 | 40.9% | 39.2% | 16.4% | Best published harness; see note above |
| Coding Agent Index-AA | 46 | 37 | 31 | Artificial Analysis |
| FrontierCode 2.0-AA, mergeability | 27.2% | 34.9% | 25.2% | Artificial Analysis; Haiku leads |
| GDP.pdf-AA, professional work over long documents | 22.8% | 20.8% | 22.8% | Artificial Analysis |
| OSWorld 2.1, computer use (partial / binary) | 69.5% / 33.8% | 68.9% / 27.8% | 50.5% / 17.0% | Meta internal; partial-credit headline |
| PerceptionBench, with tools | 80.6% | 78.6% | 62.8% | Meta internal |
| Artificial Analysis Intelligence Index | 36 | 36 | 28 | Artificial Analysis |
The sharpest lead in the independent rows is AutomationBench-AA, a strict pass@1 score on simulated business workflows: Ember 37.8 percent, Luna 20.2 percent, Haiku 13.9 percent. VentureBeat noted a second Artificial Analysis figure for the same suite, 68.1 percent, which awards credit for individual objectives completed without breaking the rules. The two numbers answer different questions. If a workflow only counts when every step succeeds, the 37.8 percent figure is the one that matters.
Coding is more mixed, and Meta’s own table shows it. Ember leads the Coding Agent Index and, in Meta’s harness comparison, Terminal-Bench 4.0. Haiku leads FrontierCode 2.0, which judges whether engineering changes are ready to merge, 34.9 percent to Ember’s 27.2 percent. On OSWorld 2.1, Ember’s 69.5 percent partial-credit score falls to 33.8 percent when tasks are graded all or nothing. VentureBeat’s reading is a fair one: the results make a stronger case for giving Ember a bounded job with a clear check than for handing it an entire engineering ticket or a desktop workflow unattended. That is also a useful contrast with Anthropic’s same-day Fable 5.2 and Mythos 5.2 launch, which the company positioned at the other end of the stack — expensive frontier models for the hardest technical work, not a high-volume execution tier.
Availability, multimodal inputs, and how Ember was trained
Meta said Muse Ember is available today on Muse Code and the Meta Model API as muse-ember-1.0, and that it supports the Responses API. The launch post also said Muse Ember, Muse Spark, and Muse Wonder “are now available on Google Cloud and Oracle Cloud.” VentureBeat reported that the model is offered through Responses, Chat Completions, and Messages endpoints; that reasoning effort ranges from minimal to max, with max as the default; and that Google Cloud’s preview documents text and image inputs. We have not independently confirmed every cloud SKU or default. Parameter counts and weights were not published. Ember is a proprietary hosted model, not an open-weight release like Glimmer or Spark Open.
Meta tied the model to a promise it made when it introduced Muse Wonder and Muse Wonder Full: that it would use those larger systems to develop specialized models it could bring to a wider audience. “Muse Ember is the first trained using Muse Wonder Full,” the company wrote. Its intended split is Muse Wonder for the hardest problems, Muse Spark for coding and agentic workflows, and Muse Ember for repeatable coding and agentic tasks at scale. The launch post also published selected early-tester comments from Box, Cline, Oracle, Cloudflare, and You.com — including Cloudflare’s claim that Ember ran up to 30 percent faster at the tail than the comparable model it tested, and You.com’s claim of a 2.2-second p95 with fewer output tokens. Those are partner remarks chosen by Meta, not independent audits.
Safety claims, and what Ember is not
Meta said it evaluated Ember under its Superintelligence Scaling Framework, covering chemical and biological, cybersecurity, and loss-of-control risks, and that “across the categories assessed, our evaluations found that Muse Ember remained below the framework’s high-risk thresholds.” It also said the small model received the same core safety training and testing as larger ones, and that it was tested on whether it respects user intent, handles sensitive information, resists hidden instructions in web pages, documents, or tool outputs, and seeks human input before high-impact actions. Those are Meta’s findings. The company did not publish the underlying evaluation data in the launch post.
A few boundaries are worth keeping straight. Ember is not a consumer Muse feature like Spaces or realtime voice. It is not, on the public record, a replacement for Spark in every coding workflow, and it is not ahead of Haiku on merge-ready engineering changes. It is not an open model you can download and run yourself. And the $0.10 / $0.50 price is not the rate for every prompt: once a request crosses 200,000 tokens, the bill jumps. For teams that need to keep inference on their own machines, Meta’s open-weight options remain Glimmer and Spark Open, which are a different product and a different cost structure.
FAQ
What is Muse Ember?
Muse Ember is Meta Superintelligence Labs’ November 9, 2026 release of its smallest and most efficient Muse model. Meta designed it for high-volume agent work — tool calls, coding subtasks, routing, extraction, and long-document parsing — and said it should be paired with a larger model such as Muse Wonder, which plans while Ember handles the volume.
How much does Muse Ember cost?
For prompts up to 200,000 tokens, Meta lists $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens. Above that cutoff, the rates are $0.30, $0.012, and $1.50, and VentureBeat reported that the higher rates apply to the entire request. Meta’s launch comparison puts cost per Artificial Analysis Intelligence Index task at about $0.14 at maximum effort.
Does it really match Claude Haiku 5.5 at one-third the cost?
On Artificial Analysis’s Intelligence Index, as published by Meta on launch day, Ember at max scored 35.8 (rounded to 36) at an estimated $0.136 per task, and Haiku at max scored 36.2 (also rounded to 36) at $0.411. That is the comparison Meta summarized as matching Haiku’s best score at about a third of the cost. It is a benchmark-cost comparison at list prices, not a guaranteed production bill, and it does not mean Ember wins every benchmark: Haiku leads FrontierCode 2.0.
Is Muse Ember better at coding than Haiku or Luna?
It depends on the test. Ember led Artificial Analysis’s Coding Agent Index (46 versus 37 and 31) and Meta’s Terminal-Bench 4.0 comparison, which used different harnesses for each model. Haiku led FrontierCode 2.0, 34.9 percent to 27.2 percent, on tasks judged by whether changes are ready to merge. Treat “better at coding” as task-specific, not as a single ranking.
Where can developers use it?
Meta said Ember is available now on Muse Code and the Meta Model API as muse-ember-1.0, with a 1 million-token context window and up to 512,000 output tokens, and that Ember, Spark, and Wonder are available on Google Cloud and Oracle Cloud. VentureBeat reported Responses, Chat Completions, and Messages endpoints, with max reasoning as the default. Weights and parameter counts were not released.
How does Ember relate to Muse Spark and Muse Wonder?
Meta’s stated portfolio is Wonder for the hardest problems, Spark for coding and agentic workflows, and Ember for repeatable tasks at scale. On the Intelligence Index, Ember scored 36 at about $0.14 per task versus Spark 1.3 at 37 and $2.39. Meta said Ember is the first model it trained using Muse Wonder Full.
Are Ember’s weights open?
No. Ember is a proprietary hosted model. Meta’s downloadable Muse options, from earlier releases, are Glimmer and Spark Open, which are separate models with their own licenses and cost profiles.
What did Meta say about safety?
Meta said evaluations under its Superintelligence Scaling Framework found Ember remained below high-risk thresholds in the chemical and biological, cybersecurity, and loss-of-control categories it assessed, and that it tested the model’s behavior when acting on a user’s behalf. Those claims come from Meta’s launch post; the underlying data was not published there.