Muse Ember Scores 36 on Artificial Analysis, Independently: Is It Actually Cheaper Per Finished Task?

Promise Ledger: Ember

Promise Ledger

A standing box for this launch, placed here so the verdict sits next to the headline. Last updated November 10, 2026, from Meta's November 9 research post and the Artificial Analysis article of November 9, both read in full on November 10, and the November 10 distribution checks listed in Sources.

Meta claimed, November 9: Ember matches Haiku 5.5's best Artificial Analysis score at about a third of the cost per task, beats GPT-6 Luna at every reasoning setting, and sits on the efficiency frontier, with no model scoring higher for less per task.
Independently checked, Artificial Analysis, November 9: the score is real in its own published table. Ember at max reasoning scores 36, Haiku 5.5 at max reasoning scores 36, and Luna at max reasoning scores 28. Cost per Intelligence Index task: Ember $0.14, Haiku 5.5 $0.41, Luna $0.15.
The setting caveat: Ember at xhigh reasoning scores 34 and costs $0.12 per task, which Artificial Analysis calls the cheapest model scoring 34 or higher. The $0.12 figure that circulated in launch coverage belongs to Ember at xhigh. It is not Luna's price, and it does not buy a score of 36.
A correction this post makes on the record: early summaries of the Artificial Analysis result, including this site's own November 10 news file, carried a Luna tie at 36 and a Luna cost of $0.12 per task. The article itself, read in full on November 10, says neither. Luna is 28 at $0.15. This post uses the article.
Still open, with one leg now partly answered: a first developer reaction surfaced on November 10, in eesel AI's review of that date, read in full that day. It reports Simon Willison's pelican test passed, in his words "the pelican was fine," and a Hacker News launch thread at 228 points. That is a secondary review reporting two early signals. It is not a review of Ember on a real workload, and this post does not treat it as one. Also still open: any statement about the consumer Muse app, and a native Oracle Cloud page for Ember. Meta's November 9 post says Ember, Spark and Wonder are available on Google Cloud and Oracle Cloud. Oracle's own pretrained models page, checked November 10, lists no Muse model.
Unresolved on price: eesel AI's November 10 review, read in full, reports tiered pricing for Ember, $0.10 and $0.50 per million tokens on prompts up to 200K tokens, then $0.30 and $1.50 per million above 200K. The tier was not in the text of Meta's November 9 post read in full on November 10, and no Meta page checked that day confirms it. This post prints the tier as reported by eesel AI, labeled secondary and unresolved, not as Meta's stated price.
Next recheck, a promise with a date: this site will recheck developer reaction, the consumer app question, and the Oracle page on November 17, 2026, and will publish the result on that date even if all three are still missing.

The short version: the independent table backs the headline, at max reasoning

Meta's launch post on November 9 made the claim in one sentence: Ember delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting. Launch claims that cite a benchmark the vendor did not run are worth checking against the benchmark owner, and that check now exists. Artificial Analysis published its own article the same day, titled "Meta's Muse Ember is the cheapest model scoring 36 on the Intelligence Index," and its table is the source for every score in this post.

The table settles two things and complicates one. It settles the score: 36 for Ember at max reasoning, the same as Haiku 5.5 at max reasoning. It settles the cost at that score: $0.14 per task against $0.41, which is 34 percent of Haiku's cost, close to the third Meta claimed. The complication is Luna. Meta's wording, that Ember beats every Luna setting, holds in the Artificial Analysis table, but not by a tie. Luna at max reasoning is eight points behind, at 28, and costs a cent more per task than Ember at max. Anyone who read a summary saying Luna also scored 36 at $0.12 was reading a garbled version of the Ember xhigh line, and the ledger above corrects it with the source named.

What Artificial Analysis measured

The score is the Artificial Analysis Intelligence Index, version 6.0.1 in the November 9 article. The article describes the index as built from nine evaluations: AA-Briefcase v1.1, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode 2.0, AA-Omniscience v1.1, GDPval, HLE-Diamond, CritPt v2 (Hard Subset), and Terminal-Bench-Science 0.1.

That list matters because 36 is not a grade on one quiz. It is a composite across agent work, terminal coding, engineering tasks, and hard knowledge questions, run by a firm that is not Meta. It is still one firm's composite, with its own choice of tasks and weights, and a score of 36 on this index does not predict what Ember will do on your workload. It does answer the narrower launch question honestly: an outside tester, publishing under its own name, got the score Meta announced.

Two things the Artificial Analysis article does not contain are worth naming, because launch coverage attached them to it. The article states no context window figure, and it contains no Vals AI score for Ember. The 1M token context figure in this post comes from Meta's own post, labeled as Meta's. A Vals AI figure of 60.48 percent appeared in summaries this site filed on November 10; it is not in the article read that day, so this post does not print it as an Artificial Analysis result.

Cost per task is not cost per million tokens

Here is the part that decides whether Ember saves a developer money, and it is easy to misread. On list price, Ember, Haiku 5.5 and Luna charge the same input and output prices in the Artificial Analysis comparison: $0.10 per million input tokens and $0.50 per million output tokens. If list price were the whole story, the three models would cost the same to run.

They do not, for two reasons the article gives. First, cached input: Ember charges $0.004 per million cached input tokens, against $0.01 for Haiku 5.5 and Luna. Artificial Analysis calls that 60 percent cheaper, and cached input is the price that matters for agent work that resends a long prompt, a codebase, or a document set on call after call. Second, cost per task is a measured figure, not a list price. It reflects how much work, in tokens, a model spends to finish the index tasks at a given reasoning setting. Ember at max reasoning finishes at $0.14 per task. Turn the reasoning down to xhigh and it finishes at $0.12, with the score falling from 36 to 34.

Field note

The practical reading is plain. If you buy tokens, Ember looks tied with its two rivals until caching enters the bill. If you buy finished tasks, which is what an agent workload actually consumes, Ember at max reasoning costs about a third of Haiku 5.5 per task at the same index score. Price your own workload the second way: run the same set of tasks on both models, count finished tasks, and divide the bill. A per million token comparison will tell you the wrong answer for this model.

Caution

Per task cost moves with the reasoning setting, and the cheapest line in the table is not the headline score. Ember at $0.12 per task is the xhigh setting, scoring 34. Ember at 36 costs $0.14. Any quote of Ember pricing that pairs $0.12 with a score of 36 has mixed two rows of the same table, which is exactly the error the ledger at the top of this post corrects.

Meta's own table, including the row Ember loses

Meta's November 9 post publishes a wider set of results, credited to Artificial Analysis, dated November 9, 2026, with some rows marked as Meta internal evaluations. These are Meta presented figures, and this post labels them that way. A few carry the argument, and one cuts against it.

On AutomationBench-AA, end to end business workflows, Meta's table gives Ember 37.8, Haiku 5.5 at 13.9, and Luna at 20.2. On Agents' Last Exam, end to end agent tasks, Ember scores 61.6, against 57.0 for Haiku 5.5 and 57.1 for Luna. On Terminal-Bench 4.0, agentic terminal coding, Ember scores 40.9, against 39.2 and 16.4. On the Coding Agent Index-AA, Ember scores 46, against 37 and 31.

The adverse row is FrontierCode 2.0-AA, real engineering tasks judged on mergeability: Haiku 5.5 scores 34.9, Ember scores 27.2, and Luna scores 25.2. On the task closest to shipping code a reviewer would merge, the model Ember ties on the headline index beats it by nearly eight points. Meta also reports a Personal Superintelligence Index score of 46.8 for Ember, against 36.1 and 37.0, and that row is marked as Meta's internal composite index. It gets the weight an internal index deserves: printed, labeled, and not used in this post's verdict, which rests on the Artificial Analysis article itself.

What Ember is, and where you can use it

Muse Ember, model name muse-ember-1.0, is the smallest model in the Muse family, built for the repeat work agents do all day: tool calls, coding subtasks, routing, extraction, and long document parsing, in Meta's own description. It is the first model trained using Muse Wonder Full. It natively takes text, image, audio and video input. Meta gives it a 1M token context window and up to 512K output tokens, with Responses API support.

Availability, in Meta's words on November 9: Muse Code and the Meta Model API, at $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens. One price question is unresolved. Eesel AI's November 10 review, read in full that day, reports a longer prompt tier above 200K tokens, at $0.30 per million input and $1.50 per million output. That tier was not in the post text read in full on November 10, and no Meta page checked that day confirms it, so the tier above is printed as eesel AI's report, not as Meta's price. The same review reports no batch API and limits of about 3,000 requests per minute and 4 million tokens per minute. Those two figures are also secondary only in this post.

The boundary that matters for most readers of this site is unchanged. No consumer Muse app availability, and no timeline for it, was stated in the November 9 announcement. If you use Muse as an app, Ember changes nothing in your account today. If you build with Muse Code or the API, it is a priced option you can call now, and the Artificial Analysis table above is the check on what it is worth.

The distribution sidebar: where else Ember showed up

A separate question trailed the launch, tracked as the next named check on this site: which gateways and clouds actually list Ember, verified on the operator's own page rather than in launch roundups. By the morning of November 10, that check had run on most legs. Vercel's changelog confirms Ember live on AI Gateway as meta/muse-ember. Netlify's changelog confirms Muse Ember 1.0 live in its AI Gateway and Agent Runners at the same prices. Cloudflare's model page confirms Ember live on Workers AI at the same prices. Google Cloud's partner models documentation confirms Muse Ember 1.0 with a November 9, 2026 release date, in Preview, with a 1M token input context and function calling.

Oracle is the one leg that stays open. Meta's November 9 post says Ember, Spark and Wonder are now available on Google Cloud and Oracle Cloud. Oracle's own page, read on November 10, lists no Muse model. Meta stated is not the same as operator confirmed, so Oracle stays in the still open row of the ledger, with the November 17 recheck date on it.

The other November 9 launch: Muse on Apple Watch and Wear OS

Ember was not the only Muse launch dated November 9. Meta's @Muse account on X announced that day, November 9 at 4:25 PM in the post read on November 10, that Muse is available on Apple Watch and Wear OS. TechCrunch reported the launch on November 10, in a piece by Aisha Malik read in full that day: a user can interact by voice calls, text messages, or dictation, with a personalized avatar, and the watch app requires a paired phone and a Muse account. The same piece puts the launch about a month after the dedicated iPad app, and repeats that Muse is one of 12 apps to pass 10 million U.S. downloads in its first 40 days.

Two limits belong in the same paragraph as the announcement. The Gadgeteer, on November 10, read in full, lists requirements of iOS 18 or later, watchOS 9 or later, and Android 10 or later, with an 18 plus gate, and lists notifications, shortcuts and a watch face, and tasks and connectors from the wrist. The Gadgeteer is a secondary source, not a Meta page, and its country list omits Mexico, so its list is incomplete. Its wording that the Wear OS app is waiting in the Google Play Store also suggests Wear OS availability may lag the announcement. That suggestion is secondary too. No source read on November 10 stated the plan or token treatment for watch use, and no source stated whether the watch works on cellular away from the paired phone. Both stay open here, and neither is guessed at.

For a reader of this site, the split on November 9 is now plain. Ember is a developer model with no stated consumer app rollout. The watch app is the consumer launch of the same two days, and it changes nothing about Ember's price or score.

What did not change on November 9

Consumer pricing did not. US pricing for the Muse app is still free with a usage limit, Power at $20 a month, and Maximum at $100 a month, and Ember is not sold as an app tier. The country list did not change either: Muse is available in the US, Canada, Japan, Singapore and Mexico, on this site's dated record. The UK has been reported as coming within weeks, with no date, partner or GBP pricing published. India has published INR pricing, reported at 1,799 rupees a month for Power and 8,999 rupees for Maximum, with the launch date itself unclear and the service waitlisted there, per The Hindu on November 9. Priced and waitlisted is not launched, and this post keeps India off the live list. Spaces sharing also remains not live: Meta's November 6 post describes it as exploring ways to share, in its own words.

Next: what does Ember cost on your tasks, not on an index?

The November 17 recheck covers the open legs in the ledger: developer reaction on real workloads, beyond the pelican test and thread count reported on November 10, a consumer app statement, Oracle's native page, and a Meta page confirming or dropping the tiered price above 200K tokens. The reader question for this post is the one an index cannot answer for you: if you run agent busywork, what does Ember cost you per finished task, not per million tokens? Name the task set. Name the reasoning setting. Name the bill for Ember and for the model it replaced. Answers in that form will shape the follow up more than another benchmark table will. The xhigh and max rows in this post show why: the answer moves when the setting changes.

Quick answers

What did Artificial Analysis score Muse Ember?

36 on its Intelligence Index, version 6.0.1, at max reasoning, in its November 9, 2026 article. That ties Claude Haiku 5.5 at 36. GPT-6 Luna at max reasoning scores 28.

Is Ember actually the cheapest model at that score?

Yes, in the Artificial Analysis table. Ember at max reasoning costs $0.14 per Intelligence Index task, against $0.41 for Haiku 5.5, and Artificial Analysis calls it the cheapest model scoring 36 or higher. Luna costs $0.15 per task but scores 28, so it is not a cheaper route to the same score.

Where does the $0.12 per task figure come from?

Ember itself, at xhigh reasoning. At that setting Ember scores 34 and costs $0.12 per task, which Artificial Analysis calls the cheapest model scoring 34 or higher. It is not Luna's cost, and it does not come with a score of 36.

What does Ember cost in tokens?

Meta's list pricing, November 9, 2026: $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens. Haiku 5.5 and Luna charge the same input and output prices in the Artificial Analysis comparison, with cached input at $0.01 per million. Eesel AI, November 10, reports a higher tier above 200K tokens, at $0.30 input and $1.50 output per million. No Meta page checked on November 10 confirms that tier, so it is unresolved in this post.

Can I use Ember in the Muse app?

No statement says so. Meta announced Ember on Muse Code and the Meta Model API as muse-ember-1.0 on November 9, 2026. No consumer app rollout or timeline was stated.

Did Muse launch anything for app users on November 9?

Yes, a different launch. Meta's @Muse account announced Muse on Apple Watch and Wear OS on November 9, 2026. TechCrunch, November 10, reports voice calls, text messages and dictation, with a personalized avatar, and that the watch app requires a paired phone and a Muse account. Plan and token treatment for watch use was not stated in any source read on November 10.

What is Ember's context window?

1M tokens, with up to 512K output tokens, per Meta's November 9 post. The Artificial Analysis article itself states no context figure.

Does Ember beat Haiku 5.5 at everything?

No. In Meta's own table, Haiku 5.5 scores 34.9 on FrontierCode 2.0-AA against Ember at 27.2, while Ember leads on AutomationBench-AA, Agents' Last Exam, Terminal-Bench 4.0 and the Coding Agent Index-AA in the same table.

Frequently Asked Questions

Q: Is this an independent test, or Meta grading its own launch?

A: The headline figures in this post, the score of 36, the tie with Haiku 5.5, Luna at 28, and the cost per task figures, come from Artificial Analysis's own November 9, 2026 article, read in full on November 10. Meta's post also publishes a wider table credited to Artificial Analysis, plus rows marked as Meta internal evaluations. This post keeps those two layers apart and rests its verdict on the Artificial Analysis article.

Q: Why did other summaries say Luna tied Ember at 36 and cost $0.12?

A: Those summaries mixed rows. In the article, Luna at max reasoning scores 28 and costs $0.15 per task. The $0.12 figure is Ember at xhigh reasoning, scoring 34. This site's own November 10 news file carried the mixed version before the article was read in full for this post, and the Promise Ledger at the top corrects it on the record.

Q: If the token prices are the same, why is Ember cheaper per task?

A: Two reasons stated in the sources. Cached input is cheaper on Ember, $0.004 per million against $0.01 for Haiku 5.5 and Luna, a 60 percent difference per Artificial Analysis. And cost per task is measured from the work a model spends finishing a task at a setting, not from list price alone, which is why the same list prices produce $0.14, $0.41 and $0.15 per task in the table.

Q: Which reasoning setting should a developer pick?

A: The table gives a clean trade. Max reasoning buys a score of 36 at $0.14 per task. Xhigh buys 34 at $0.12. Which one is right depends on whether the two index points are worth two cents a task on your workload, and only a run on your own tasks answers that. This post names the trade; it does not pick for you.

Q: Where else can I run Ember besides Muse Code and the API?

A: On operator pages checked November 10, 2026: Vercel AI Gateway, Netlify AI Gateway and Agent Runners, Cloudflare Workers AI, and Google Cloud, where the partner models documentation lists Muse Ember 1.0 in Preview with a November 9 release date. Oracle Cloud is stated by Meta but was not found on Oracle's own models page on November 10, so it stays labeled Meta stated, not operator confirmed.

Q: Did Ember change Muse app prices or countries?

A: No change was found. App pricing in the US is free with a usage limit, Power at $20 a month, and Maximum at $100 a month. Muse is live in the US, Canada, Japan, Singapore and Mexico. The UK is reported within weeks with no date, and India is priced and waitlisted, not launched, per The Hindu on November 9, 2026.

Q: What happens on November 17?

A: This site rechecks developer reaction to Ember on real workloads, any Meta statement about the consumer Muse app, Oracle's native page for Ember, and a Meta page confirming or dropping the tiered price above 200K tokens reported by eesel AI on November 10. The result will be published on November 17, 2026 even if those legs are still missing, the same dated promise used in the Promise Ledger at the top of this post.

Sources

Artificial Analysis, "Meta's Muse Ember is the cheapest model scoring 36 on the Intelligence Index," November 9, 2026, artificialanalysis.ai. Read in full on November 10, 2026. The independent figures used for this post's verdict: Intelligence Index version 6.0.1, built from nine named evaluations; Ember at max reasoning scores 36 at $0.14 per task, Haiku 5.5 at max reasoning scores 36 at $0.41 per task, Luna at max reasoning scores 28 at $0.15 per task, and Ember at xhigh scores 34 at $0.12 per task, the cheapest model scoring 34 or higher. Token comparison from the same article: Ember, Haiku 5.5 and Luna share $0.10 input and $0.50 output per million tokens, with cached input at $0.004 for Ember against $0.01 for the other two, described as 60 percent cheaper. The article states no context window and contains no Vals AI score.
Meta AI Research, "Introducing Our Smallest, Most Efficient Model: Muse Ember," research.meta.ai, November 9, 2026. Read firsthand on November 10, 2026. The launch facts: smallest and most efficient model, built for tool calls, coding subtasks, routing, extraction and long document parsing; muse-ember-1.0 on Muse Code and the Meta Model API; $0.10 per million input, $0.004 per million cached input and $0.50 per million output; 1M token context and up to 512K output; Responses API support; native text, image, audio and video input; first model trained using Muse Wonder Full. The wider results table, credited to Artificial Analysis, November 9, 2026, with internal rows marked: AutomationBench-AA 37.8 against 13.9 and 20.2, Agents' Last Exam 61.6 against 57.0 and 57.1, Terminal-Bench 4.0 at 40.9 against 39.2 and 16.4, Coding Agent Index-AA 46 against 37 and 31, FrontierCode 2.0-AA 27.2 against Haiku 5.5 at 34.9 and Luna at 25.2, and a Personal Superintelligence Index of 46.8 marked as Meta's internal composite. The closing availability statement for Google Cloud and Oracle Cloud is also from this post. No consumer app rollout, and no longer prompt pricing tier, appeared in the post text read on November 10.
Distribution checks, November 10, 2026, from this site's news watch of that date: Vercel's changelog of November 9 confirms Ember on AI Gateway as meta/muse-ember; Netlify's changelog of November 9 confirms Muse Ember 1.0 in AI Gateway and Agent Runners at the launch prices; Cloudflare's Ember model page confirms it live on Workers AI at the launch prices; Google Cloud's partner models documentation confirms Muse Ember 1.0, release date November 9, 2026, in Preview, with a 1M token input context and function calling; Oracle's pretrained models page, read the same day, lists no Muse model.
Eesel AI, "Meta Muse Ember review (2026)," November 10, 2026, read in full on November 10, 2026 in this site's demand log sweeps of that date. Secondary source, not a Meta page. The figures taken from it, labeled secondary in the text: tiered pricing of $0.10 and $0.50 per million tokens on prompts up to 200K tokens, then $0.30 and $1.50 per million above 200K, cached input at $0.004 per million, no batch API, and limits of about 3,000 requests per minute and 4 million tokens per minute; and the first developer reaction reported for Ember, Simon Willison's pelican test, in his words "the pelican was fine," and a Hacker News launch thread at 228 points.
Apple Watch and Wear OS checks, November 10, 2026: Meta's @Muse account on X, November 9, 2026, 4:25 PM, read in full on November 10, announcing Muse on Apple Watch and Wear OS. TechCrunch, November 10, 2026, Aisha Malik, read in full: voice calls, text messages and dictation, a personalized avatar, a paired phone and a Muse account required, the launch about a month after the dedicated iPad app, and Muse as one of 12 apps to pass 10 million U.S. downloads in its first 40 days. The Gadgeteer, November 10, 2026, read in full, secondary source: requirements of iOS 18 or later, watchOS 9 or later and Android 10 or later, an 18 plus gate, notifications, shortcuts and a watch face, and tasks and connectors from the wrist; its wording suggests the Wear OS app may lag in the Google Play Store, and its country list omits Mexico. No source read on November 10 stated plan or token treatment for watch use, or whether the watch works on cellular away from the paired phone.
Editorial checks, November 9 and 10, 2026: standing facts held flat from this site's November record, US app pricing at free with a usage limit, Power at $20 a month and Maximum at $100 a month; Muse live in the US, Canada, Japan, Singapore and Mexico; UK reported within weeks with no date, partner or GBP pricing; India priced at 1,799 rupees for Power and 8,999 rupees for Maximum with the launch date unclear, per The Hindu, November 9, 2026; Spaces sharing still described by Meta on November 6 as exploring ways to share, not live.