Meta Releases Muse Ember for Repeat Agent Work at Ten Cents per Million Input Tokens

On 9 November 2026 Meta posted Muse Ember on the Model API at $0.10 per million input tokens. The research note and evaluation report name the scores, the harnesses, and the tasks where a larger model still leads.

PUBLISHED //
DURATION // 25 MIN READ
A dark walnut desk at night with a small warm point of light on glass and thin blue arcs above a closed notebook, no text
FIG.01 // VISUAL DOCUMENTATION A dark walnut desk at night with a small warm point of light on glass and thin blue arcs above a closed notebook, no text

Meta posted Muse Ember on 9 November 2026. The third note in the thread, time-stamped 18:04 UTC, states a price and a destination: the model is available that day on the Meta Model API at $0.10 per million input tokens and $0.50 per million output tokens, and the short link expands to go.meta.me/muse-ember. That address redirects to the research post Introducing Our Smallest, Most Efficient Model: Muse Ember, dated 9 November 2026 on the page and in its schema markup. A methodology PDF linked from that post, the Muse Ember evaluation report, spells out harnesses and scoring rules the charts assume.

The thread is shorter than the blog. The opening post calls Muse Ember the smallest and most efficient model, built for complex reasoning execution and everyday agent work. The next sentences say it can carry workflows across apps, call tools in sequence, run tests, and make targeted code fixes. The written text then stops at “With a 1M-token context window.” The only expanded URL on that status is a photo. The photo is not a source for the numbers below. The second post, one second later, says efficiency matters for agents that perform hundreds of small tasks, and that on the Artificial Analysis Intelligence Index the model achieves performance comparable to Claude Haiku 5.5 at roughly one-third of the cost. That status also ends on a photo. The third post is the one with a page behind it.

A fourth post, at 21:58 UTC the same day, says Meta is also bringing Muse Ember to Google Cloud along with Muse Wonder and Muse Spark 1.3. It does not repeat the price. It does not mention Oracle. The research post’s closing line is wider: Muse Ember, Muse Spark, and Muse Wonder are now available on Google Cloud and Oracle Cloud. The Spark version is specified in the Google Cloud post and in a comparison table on the blog. The closing sentence says Muse Spark without a version number. Those are three different sentences.

What the research post says the model is for

The blog’s first paragraph is more specific than the opening tweet. Muse Ember is described as the smallest and most efficient model, designed for complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing. Later, under Agents, the post says agents spend most of their tokens on many small, structured steps, and that Ember can carry a business workflow end to end, pulling records, calling tools in sequence, and updating systems across apps. Under Coding, it says the model is built for the coding subtasks an agent sends off hundreds of times a session: reading files, running tests, making targeted fixes.

That job description is the right frame for the scores that follow. A model aimed at repeated subtasks can post a strong number on a workflow bench and a weaker number on a bench that asks whether a maintainer would merge a change. The page contains both outcomes. Reading only the first is a choice, not a summary.

The context line the tweet left unfinished is completed on the blog. With a 1M-token context window and up to 512K output tokens, the model can take in a full codebase or a document set in a single call, the post says. The API name is muse-ember-1.0. It is available on Muse Code and the Meta Model API, and it supports the Responses API. The post does not define the Responses API in this note, so the support claim is a name, not a feature list.

Multimodal coverage is also blog text, not thread text. Ember 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. The sentence is a capability claim. It is not a measured error rate for OCR, transcription, or video. Those rates, where the page has them, sit in later tables.

“Smallest” is Meta’s ranking inside its own family. No parameter count appears in the thread, the blog, or the methodology PDF. A buyer who needs weights, active parameters, or a distillation diagram will not find them in this launch set. The efficiency claim has to be read as a price-and-score claim, which is what the charts actually print, rather than as a claim about model size in the engineering sense.

Two prices sit under the headline rate

The tweet’s price and the blog’s price match on the two headline numbers: $0.10 per million input tokens and $0.50 per million output tokens. The blog adds a cached input rate of $0.004 per million tokens, and it limits that card to prompts up to 200K tokens. For longer prompts, the same footnote sets Muse Ember at $0.30 per million input tokens, $0.012 per million cached input tokens, and $1.50 per million output tokens.

The 1M-token window and the 200K pricing break are easy to paste into the same slide and then forget. A call that stays under 200K tokens is on the headline card. A call that uses the window the announcement is proud of, a full codebase or a long document set, crosses into the higher card if the prompt is over 200K tokens. The higher input rate is three times the headline input rate. The higher output rate is three times the headline output rate. The higher cached rate is three times $0.004. None of that is hidden. It is in a footnote under the charts, which is where a comparison deck tends to drop it.

Cached input at $0.004 is 25 times cheaper than uncached input at $0.10, on the short-prompt card. For an agent loop that resends a stable prefix, the cached rate is the number that decides whether “hundreds of small tasks” stays cheap. The page states the rate. It does not state a cache-hit policy, a minimum prefix length, or how Muse Code and the Model API decide what counts as cached. A cost model that assumes every repeated token bills at $0.004 is assuming a rule the note does not give.

There is a second way the headline can mislead a budget. Output is five times the short-prompt input price, $0.50 against $0.10. An agent that “makes targeted code fixes” and writes long patches is an output-heavy call. The input discount does not apply to those tokens. On the long-prompt card the output rate is $1.50 per million, still five times the matching input rate. A routing policy that sends volume to Ember because input is cheap, then lets the model narrate every file it touched, can spend the savings on output. The page gives the rates. It does not give a typical output-to-input ratio for the agent jobs it names.

The comparison table later prints a different cost figure: about $0.14 per Artificial Analysis task for Ember at max reasoning, against $2.39 for Muse Spark 1.3 at max. That $0.14 is a task cost, not a token price. Dividing $2.39 by $0.14 gives about 17. The intelligence index on that same table is 36 for Ember and 37 for Spark 1.3. Token prices on the table are $0.10 input and $0.50 output for Ember, against $1.25 input and $4.25 output for Spark 1.3. The page’s own argument for keeping a larger model in the portfolio is sitting in that row: one index point, and a task cost that is not in the same range.

The scatter chart’s Ember points are not $0.14. They are $0.116 and $0.136. The comparison table’s $0.14 is the max-effort index-task figure printed later. A reader who averages them, or who treats $0.14 as the price of every setting, is mixing two captions. One chart caption calls Ember’s costs estimates at proposed pricing. A later line on the same chart says cost at list pricing. The page footnote says cost per task uses each model’s list pricing. “Proposed” and “list” are different words, printed on the same figure. The API section states list prices. The chart should be read next to that tension.

What “about a third of the cost” is doing

The second tweet says performance comparable to Claude Haiku 5.5 at roughly one-third of the cost, on the Artificial Analysis Intelligence Index. The blog tightens the wording. Ember delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting: it matches Haiku’s best score at about a third of the cost and beats every Luna setting. It also sits on the efficiency frontier of the full Artificial Analysis leaderboard: no model scores higher for less per task.

The scatter lists points, and the caption says they come from raw Artificial Analysis results dated 9 November 2026. The listed Ember points are a score of 33.9 at $0.116 per task and a score of 35.8 at $0.136 per task. Haiku’s listed points run from 24.6 at $0.092, through 29.6 at $0.159 and 34.3 at $0.248, up to 36.2 at $0.411. Luna’s listed points are 21.4 at $0.073, 23.3 at $0.104, and 27.8 at $0.152. The page also prints reasoning-effort labels beside the chart, including medium, high, xhigh, and max. The label order in the page markup does not lock each label to a point tightly enough to invent a mapping. The prose claim does not need that lock.

Haiku’s highest listed score is 36.2 at $0.411. A third of $0.411 is $0.137. Ember’s $0.136 point, with a score of 35.8, is the point the “about a third” sentence is standing on. The scores are close. They are not printed as the same number. A later table prints an Artificial Analysis Intelligence Index of 36 for Ember at max. Spark 1.3 at max is 37 on that table. The scatter’s Ember high point is 35.8, not 36. The page does not explain the tenth. Treating 35.8 and 36 as one measurement is a reasonable reading of rounding. Treating them as two independent wins is not.

“Beats every Luna setting” fits the listed Luna scores, whose highest printed point is 27.8, under both Ember points. “No model scores higher for less per task” is a claim about the full leaderboard, not only about Haiku and Luna. The chart on the page shows three series. The sentence claims a frontier against models that are not drawn. That is Meta’s ranking of an external board. It is not a table of those other models in this note. A procurement note that repeats the frontier sentence should say it is Meta’s sentence about Artificial Analysis, and that the only series printed here are Ember, Haiku 5.5, and Luna.

The methodology report says the Intelligence Index v5.0 combines AA-Briefcase v1.1, AutomationBench, GDP.pdf, FrontierCode 2.0, Terminal-Bench 4.0, Terminal-Bench-Science 0.1, CritPt v2, AA-Omniscience v1.1, and HLE-Diamond. The result source is Artificial Analysis. A single index number is a blend. A lead on the blend can sit next to a loss on one ingredient. This page prints that loss.

The page already contains the loss

On FrontierCode 2.0-AA, at max reasoning, Claude Haiku 5.5 scores 34.9 percent, Muse Ember scores 27.2 percent, and GPT-6 Luna scores 25.2 percent. The source line is Artificial Analysis, 9 November 2026. The report says FrontierCode 2.0 comes from Cognition and grades real engineering tasks in real repositories on whether a maintainer would merge the change. Artificial Analysis runs a private 132-task set. The primary metric is accuracy.

That is the clearest counterexample on the page, and it is Meta’s page. A 7.7 point gap behind Haiku, on a mergeability exam, sits next to copy about targeted fixes and test runs. Those are related coding jobs. They are not the same grading rule. A team that routes volume to Ember because the index looks cheap, then asks it to land mergeable changes in real repositories, is asking a question the FrontierCode row already answers in Haiku’s favor on this set. Ember still leads Luna on that row, 27.2 against 25.2. The loss is specific. It is not a general collapse.

Other rows go the other way, and they should be named with their scoring rules attached.

AutomationBench-AA, at max reasoning, is 37.8 percent for Ember, 20.2 percent for Luna, and 13.9 percent for Haiku. The chart caption says strict pass@1, Artificial Analysis, 9 November 2026. The report adds the setup: Artificial Analysis’s run of Zapier’s private set of 657 business-workflow tasks across simulated applications such as Gmail, Salesforce, and Slack. Agents complete workflows through APIs. Deterministic checks inspect final state. Pass@1 here means the share of tasks completed in full, every objective met, no guardrail violated. The gap is large. The set is private. A large gap on a private set is still a large gap in the record Meta chose to print. It is not a public reproduction kit, and the launch pages do not link a dataset.

Agents’ Last Exam, at max, is 61.6 percent for Ember, 57.1 percent for Luna, and 57.0 percent for Haiku. The blog caption says this is a Meta internal evaluation on a Linux subset. The report is more precise. ALE v1.0, Linux task split, drops five of 105 tasks for every model, the ones whose input data exceeds the evaluation environment’s storage limit or whose scripts did not return a score. That leaves 100 tasks. Each task’s native script awards partial credit between 0 and 1. A few checks that use an LLM judge use GPT-5.5. Every model runs in a common internal agent loop. The official leaderboard, the report says, compares agent products such as Codex and Claude Code instead. The primary metric is the mean task score. A 61.6 to 57.1 lead is a lead on that mean, in that loop, on that Linux subset. It is not a claim about the official product leaderboard.

The Personal Superintelligence Index, at max, is 46.8 percent for Ember, 37.0 percent for Luna, and 36.1 percent for Haiku. The blog calls it an internal composite of personal and knowledge work. The footnote lists eight evaluations: Agents’ Last Exam, AutomationBench, FocusBench, JobBench, OfficeQA Pro, GDP.pdf, GDP.xlsx, and VBT. The report names OfficeQA Pro V2 and describes the pieces. JobBench is 65 professional tasks across 35 occupations, scored by a Grok 4.3 judge in the OpenCode harness, with mean reward as the metric. OfficeQA Pro V2, from Databricks, is 90 questions on U.S. Federal Accounts of Receipts and Expenditures, 1,435 documents spanning 1793 to 2024, graded deterministically with a 0 percent tolerance, in a bash sandbox with a browser tool and benchmark sites blocked. GDP.xlsx, from Surge, is 70 spreadsheet and CSV tasks in 13 domains, with Gemini 3.8 Flash grading rubrics, in an internal loop that replaces the official OpenHands loop. Verifiable Browsing Task is an internal web-browsing eval of list-style questions. FocusBench is an internal eval of personal tasks in areas such as housing, travel, IT, and finance, including work that has to continue through blockers. PSI is a Meta composite. A lead on PSI is a lead on Meta’s blend. It describes how Meta scores its own personal and knowledge-work bucket. It is not an external board.

The report’s opening rules apply to all of these internal runs. Muse Ember results are generated at max reasoning effort through the Meta Model API. Competitor runs use common settings where practical, and remain best-effort: prompts, tools, and runtime may not be optimally tuned for proprietary models. Truncated responses are graded as they are. Refusals receive zero credit and stay in the denominator rather than being dropped. Samples lost to infrastructure or grading errors are not scored. A shared harness is not the same thing as a fully tuned comparison. The PDF says so. The chart titles do not.

A harness is part of the number

Terminal-Bench 4.0, at max, prints 40.9 percent for Ember, 39.2 percent for Haiku, and 16.4 percent for Luna. The blog footnote says each model uses its best published harness: mini-swe-agent for Muse Ember, Claude Code for Claude Haiku 5.5, and Codex for GPT-6 Luna. The report goes further. Ember’s result is Artificial Analysis’s run in mini-swe-agent. Haiku’s number is Anthropic’s self-reported Claude Code result from the Haiku 5.5 system card. Luna’s number is the Codex result on the Terminal-Bench leaderboard. The bench itself is 66 harder tasks in isolated terminals, graded by executable verifiers on the final container state. The primary metric is accuracy.

A 1.7 point gap between Ember and Haiku, across three harnesses and two reporting paths, is not a same-harness result. The report’s opening rule draws this line on purpose. For controlled internal comparisons, every model uses the same benchmark-specific harness. External coding results keep the harness or agent product used by the reporting vendor, and the source is identified. Terminal-Bench on this page is the second kind. The gap to Luna, 40.9 against 16.4, is large enough that harness differences are unlikely to be the whole story, and still the page does not claim a shared harness for that row. A careful note quotes the harnesses with the percents.

The Coding Agent Index-AA is the same kind of mix. The report says version 2.0 combines Terminal-Bench 4.0, SWE-Atlas-QnA, and FrontierCode v1.1, each run inside a coding agent: Muse Code for Ember, Claude Code for Haiku, and Codex for Luna. The result source is Artificial Analysis, dated 9 November 2026 on the chart. The printed scores are 46, 37, and 31. Those numbers compare agent products as much as they compare weights. Muse Code is also a product named in the availability section, which means Ember’s index point is partly a statement about that product’s harness, not only about muse-ember-1.0 as a raw completion endpoint.

That distinction matters for anyone calling the Model API directly. The blog says Ember is available on Muse Code and on the Meta Model API. It does not say a raw API call reproduces the Coding Agent Index. The index, by the report’s definition, was run inside Muse Code. A team that wires muse-ember-1.0 into a different agent loop should expect a different number, and the launch pages do not supply that other number.

Multimodal scores depend on the credit rule

GDP.pdf-AA, at max, is 22.8 percent for Ember, 22.8 percent for Luna, and 20.8 percent for Haiku. The source line is Artificial Analysis, 9 November 2026. The report says this is Surge’s professional-work exam over supplied PDFs, in Artificial Analysis’s version: the model receives extracted text plus page images, answers without tools, and a judge checks every rubric criterion separately. The metric is all-pass, the share of answers that meet every criterion. A tie at 22.8 on an all-pass metric is a low absolute number next to the 60s and 80s elsewhere on the page. “Can work through long real-world documents” is the blog’s capability sentence. 22.8 percent all-pass is the measured sentence. Both can be true. A product brief that quotes only the first is editing the page.

OSWorld 2.1 is where the credit rule changes the headline. The blog’s summary table prints partial and binary scores: Ember 69.5 percent partial and 33.8 percent binary, Haiku 68.9 percent partial and 27.8 percent binary, Luna 50.5 percent partial and 17.0 percent binary. The chart section repeats the partial numbers and captions them as a Meta internal evaluation with partial-credit reward. The report defines the bench: OSWorld 2.1, from XLANG Lab, is the latest release of OSWorld 2.0, with 108 long-horizon computer-use workflows on a live Ubuntu desktop. The agent operates through screenshots and mouse and keyboard actions. Each task’s evaluator awards partial credit on the final state of the machine. Model-based checks are judged by Claude Sonnet 4.6. Meta runs all 108 tasks in a common internal framework, not the official runner, with each model in its provider’s computer-use configuration. The primary metric is the partial score, the mean credit per task. The strict pass rate, the share of tasks with full credit, is reported second.

The partial gap between Ember and Haiku is 0.6 points. The binary gap is wider, 33.8 against 27.8. A slide that says “69.5 on OSWorld” without the word partial is using the primary metric and dropping the strict one. The page prints both. The 69.5 figure is not a claim that the agent finished about seven in ten desktop tasks cleanly. The binary column is the clean-finish column, and for Ember it is about half the partial figure. Luna’s partial score of 50.5 against a binary score of 17.0 shows the same split in a sharper form. Partial credit is a real measurement. It is a different measurement from task completion.

PerceptionBench, with tools, at max, is 80.6 percent for Ember, 78.6 percent for Haiku, and 62.8 percent for Luna. The caption says Meta internal, with tools. The report says the bench is from Moonshot AI: 3,000 verified questions with short, uniquely determined answers, across ten capabilities, including counting, localization, depth, OCR, and perception-related hallucination. Meta’s run lets the model inspect the image in a sandbox with a bash terminal and a file viewer and editor. The official judge, gpt-oss-120b, checks each answer against the reference with the official judge prompt. The metric is pass@1. “With tools” is doing real work in that score. A no-tools perception number is not in this table. The native-understanding sentence in the blog, which says agents can read screenshots and charts without a separate model, is compatible with a tools-on exam and is not the same claim as 80.6 percent.

Partner lines are early tests with named limits

The blog prints early-testing comments from five companies. They are not the Artificial Analysis tables, and they should not be averaged into them.

Box, through Yashodha Bhavnani, VP of AI Products, says that on the Box AI Complex Work Eval, Ember shows accuracy, speed, and token efficiency. The examples given are scoring a contract against a company’s negotiation playbook, clause by clause, including an overall risk rating, and screening candidates against a detailed brief, including borderline people who still qualify as alternates. No score from that eval is printed. Cline founder Saoud Rizwan describes Ember as a fast, low-cost model for developers, and as a subagent in a harness for coding and document analysis, aimed at teams cutting cost. No latency figure is attached. Oracle’s Konstantinos Papamiltiadis, SVP of Global Partnerships, says early testing showed an encouraging ability on complex engineering tasks and strong potential for engineering teams. No task list and no score.

Cloudflare CTO Dane Knecht says speed stood out. Ember ran up to 30 percent faster at the tail than the comparable model Cloudflare tested, and it handled broad, multi-repo investigation tasks. The comparable model is not named. “Up to 30 percent” is a ceiling phrase, not a published mean, and “at the tail” is not defined as a percentile in the quote. You.com inference lead Chak Pothina says Ember completed every coding quest in their testing with the lowest latency of any model they tested: a p95 of 2.2 seconds, about 4x to 13x faster than leading open-weight models, and with 36 to 48 percent fewer output tokens. The open-weight models are not named. “Every coding quest” is their set. A p95 of 2.2 seconds is a tail latency on that set. It is not a token-price claim and not an Artificial Analysis score. The quote also ties the work to Meta Superintelligence Labs. That name appears in the quote. It is not given a separate product page in the thread.

Those sentences are usable as evidence that named teams ran early tests and were willing to be quoted on a 9 November 2026 research post. They are not usable as a ranked bake-off. The missing model names are the reason. A later slide that turns “4x to 13x faster” into a comparison against Haiku or Luna is adding a pairing the quote does not contain. You.com’s contrast is with leading open-weight models. Haiku and Luna, as named on the charts, are not identified as open-weight in this note.

Wonder plans, Ember handles volume, Spark stays in the table

One sentence on the blog is a routing policy stated without a benchmark. For the hardest problems, pair Ember with Muse Wonder. The post assigns planning to Muse Wonder and volume to Ember. Another sentence says Ember is the first model trained using Muse Wonder Full, after Meta said it would use Wonder and Wonder Full to develop specialized models it could bring to more people. That is a training-lineage claim. The note does not include a training report, a data mix, or a diagram. It is safe to repeat as Meta’s account of lineage. It is not safe to expand into a method.

The Spark comparison is the quantitative version of the same portfolio idea. Spark 1.3 at max scores 37 on the Intelligence Index against Ember’s 36, at $2.39 per index task against about $0.14, and at $1.25 per million input tokens and $4.25 per million output tokens against $0.10 and $0.50. The Google Cloud post names Spark 1.3 in the same breath as Ember and Wonder. A buyer who hears that Ember replaces Spark is not reading the table. A buyer who hears that Ember is the volume model next to Spark and Wonder is reading the sentence Meta wrote, and the price row that makes the split concrete.

Availability has the same split between a short post and a longer page. The 21:58 UTC post names Google Cloud, Ember, Wonder, and Spark 1.3. The blog adds Oracle Cloud, and it adds Muse Code plus the Model API, the model id, the Responses API, and both pricing cards. Oracle is not in the tweet. Google Cloud is in both. A status page or a cloud console was not part of this launch record, so “available today” is the claim in the posts, not a checked endpoint.

Safety is a threshold statement, not a score

Before release, Meta says it evaluated Ember under the Meta Superintelligence Scaling Framework, including chemical and biological risks, cybersecurity risks, and loss-of-control risks. Across the categories assessed, the evaluations found that Ember remained below the framework’s high-risk thresholds. The post says that even as a small model it received the same core safety training and testing. It also says Meta evaluated how Ember behaves when acting on a user’s behalf: whether it respects user intent, handles sensitive information appropriately, resists instructions hidden in web pages, documents, or tool outputs that try to redirect it, and seeks human input before high-impact actions.

No threshold value is printed. No pass rate is printed. No failure count is printed. “Below the high-risk threshold” is a classification against an internal framework, not a public score a second lab can recompute from this note. The methodology PDF covers general intelligence, agents, coding, and multimodal understanding. It does not reproduce the safety runs. A review that needs the threshold, the test set, or the refusal behavior has to ask for documents this launch post does not contain. The category names are the whole of the safety record here.

What can be repeated without adding a number

On 9 November 2026 Meta named Muse Ember, put it on the Meta Model API as muse-ember-1.0, and priced the short-prompt card at $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens, with a tripled card past 200K tokens. The context window is 1M tokens. Output runs up to 512K. The written thread states the headline input and output prices and an unfinished context clause. The research post and the evaluation report are where the rest of the numbers live, including the ones that cut against the efficiency headline.

A team writing an internal note can carry this much without inventing a bake-off. Ember is positioned as the repeat-work model in a three-model set, not as a replacement for Muse Wonder or Muse Spark 1.3. On the Artificial Analysis points Meta prints, it is cheap relative to Haiku 5.5’s best listed index point, and it trails that same Haiku on FrontierCode 2.0-AA mergeability, 27.2 percent against 34.9 percent. Terminal-Bench’s 40.9 against 39.2 is a cross-harness printout, with mini-swe-agent, Claude Code, and Codex named in the footnote. OSWorld’s 69.5 is a partial-credit mean, with a binary score of 33.8 beside it. GDP.pdf all-pass is a tie with Luna at 22.8 percent. Partner quotes name companies and roles, and they do not name the models they beat, except You.com’s unnamed open-weight set. Safety is a below-threshold statement without a published threshold.

That is a full product note. The page is more careful than a one-line reading of “smallest and most efficient,” and the careful reading is the one with the footnotes still attached. The short link in the thread is enough to reach the page. The page is not interchangeable with the thread.