Meta says Muse Ember costs a third of Haiku 5.5 per task at equal token prices; it trails on FrontierCode

A third of Haiku 5.5's cost per task on Meta's chart, at the same price per token

Meta's launch post for Muse Ember, dated 9 November 2026 by its page data, says the model "matches Haiku's best score at about a third of the cost". The nine points printed in its chart give Muse Ember 35.8 at $0.136 per task and Claude Haiku 5.5 36.2 at $0.411: a cost ratio of 0.33 (derived), with Ember 0.4 index points behind. Artificial Analysis's models page, read the same evening, lists Ember at 36.02 and $0.1359 per task and Haiku 5.5 at 36.48 and $0.4110, and lists Ember, Claude Haiku 5.5 and GPT-6 Luna at one and the same price: $0.10 per million input tokens and $0.50 per million output tokens.

With an equal price per token, a third of the cost per task is a third of the tokens. Artificial Analysis prints 84,482 output tokens per task for Ember against 225,966 for Haiku 5.5, 2.67 times fewer (derived). Meta's page prints neither number. The claim is therefore about how many tokens Ember spends on Artificial Analysis's task mix, not about Ember being cheaper per token than the others.

What that does not establish: a bill. Artificial Analysis states that each cost "is calculated from input, cache hit, cache write, reasoning, and answer token prices" on its own tasks. Meta's chart caption says "Muse Ember costs are estimates at proposed pricing", and the footer under the same chart says "Muse Ember cost at list pricing". Nothing on either page measures what Ember costs on your tasks, with your cache hit rate and your prompt sizes. Meta's own table also shows Ember behind Haiku 5.5 on FrontierCode 2.0-AA, 27.2 against 34.9 (a gap of 7.7, derived), and its pricing page bills a request over 200,000 input tokens at three times the headline rates.

Meta sells Muse Ember by the token and through Muse Code subscriptions, so its launch page is a vendor's claim, as Anthropic's and OpenAI's pages are for the models it names. Read between 20:38 and 21:00 UTC on Monday 9 November 2026. Artificial Analysis and Vals AI pages are live leaderboards and can move within the hour. No UK section follows: nothing in the pages read gives the story a UK angle.

What Meta's chart prints, and what "at any reasoning setting" covers

Meta writes that Ember delivers "more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting", and that it sits on the efficiency frontier of the full Artificial Analysis leaderboard: "no model scores higher for less per task". The chart's page text gives nine points, one per reasoning effort setting. The page prints the setting labels in a separate list, in the order xhigh, max for Ember, then medium, high, xhigh, max for Haiku 5.5, then high, xhigh, max for Luna; the pairing below follows that order (our reading), and the max points agree with Artificial Analysis's own max effort costs.

Zoom

Three of the nine points cost less per task than Ember's cheapest plotted setting ($0.116): Haiku 5.5 medium at $0.092 and Luna at high ($0.073) and xhigh ($0.104). So the claim is about the frontier, the best score reachable at each spend, not that every Ember setting beats every rival setting on both axes. Ember's xhigh point (33.9 at $0.116) does beat Haiku 5.5's high setting (29.6 at $0.159) on both axes, and every Luna point scores lower than Ember's lowest.

"Matches Haiku's best score" rests on a gap of 0.4 index points on Meta's chart and 0.46 (derived) on Artificial Analysis's page; the evaluation table prints both at 36. Meta prints no error bars. On this site's check of the 27 models on the saved Artificial Analysis page, none scores 36.02 or more at under $0.1359 per task (derived), which is consistent with the frontier sentence for that set; the page says it tracks 734 models, so the full leaderboard was not checked. The same page lists GPT-6.1 Terra (Max) at 51.33 for $0.9611 per task: 15.3 index points more for 7.1 times the cost (derived). Whether that is worth paying depends on the task, which neither chart can say.

Meta's comparison with its own larger model shows the same shape. Its page lists Muse Spark 1.3 at an index of 37 against Ember's 36, at $2.39 per task against $0.14 (17 times, derived from the printed figures). This site's earlier brief on Claude Haiku 5.5 looks at the cost per task claims Anthropic made for that model.

Meta's evaluation table: who ran each number

Meta's page prints ten rows, every model at max reasoning effort. Counting them ourselves: against Claude Haiku 5.5 Ember leads 8 rows, ties 1 and trails 1; against GPT-6 Luna it leads 9, ties 1 and trails 0. The tie with Haiku 5.5 is the Artificial Analysis Index, printed as 36 for both; the tie with Luna is GDP.pdf-AA at 22.8. The one trail is FrontierCode 2.0-AA. Of the ten rows, five are Artificial Analysis runs, four are Meta's internal evaluations, and one, Terminal-Bench 4.0, mixes three sources.

Zoom

Three cautions come from Meta's own evaluation report, a four page methodology document. First, on Terminal-Bench 4.0 the three numbers do not come from one harness. Ember's is "Artificial Analysis’s run in the mini-swe-agent harness"; Haiku 5.5's is "Anthropic’s self-reported Claude Code result" from its system card; Luna's is the Codex result on the Terminal-Bench leaderboard. Artificial Analysis's own page, read separately, gives Ember 40.9, Haiku 5.5 32.8 and Luna 12.6 on the same benchmark, so Meta's choice of Haiku 5.5's self-report (39.2) shrinks Ember's lead from 8.1 points to 1.7 (derived).

Second, the competitor numbers on Meta's internal rows are Meta's runs, and the report says they "remain best-effort" and "may not be optimally tuned for proprietary models". Agents' Last Exam excludes five of its 105 tasks for every model; OSWorld 2.1's model based checks are judged by Claude Sonnet 4.6, and PerceptionBench is judged by gpt-oss-120b. None of the internal rows carries a confidence interval, so this site cannot say whether a margin as small as 0.6 points (OSWorld) or 2.0 (PerceptionBench) is distinguishable from run to run variation.

Third, the Personal Superintelligence Index is "an internal composite of eight personal and knowledge-work evaluations", and two of its eight parts, Agents' Last Exam and AutomationBench, appear as rows of their own above. The weights are not stated, so the 10.7 point lead is not independent of two other rows. The one row Ember trails, FrontierCode 2.0-AA, is Artificial Analysis's private 132 task set judged on whether a maintainer would merge the change, and the Artificial Analysis page agrees with Meta's three figures. Meta's own developer documentation says: "for hard single-shot coding problems, knowledge-heavy questions, or visual coding, use Muse Spark or Muse Wonder rather than Muse Ember".

What the independent pages add, and where they disagree with each other

Three outside leaderboards were checked. Artificial Analysis lists Ember; Vals AI's index (version 2.2, updated 9 November 2026) lists it; the CWE-bench leaderboard does not: its results are dated 2 November 2026, before the launch, and it lists Muse Wonder but not Muse Ember. Artificial Analysis ranks Ember second of 32 on intelligence and eleventh of 32 on cost within its comparison class, and says "Meta has not disclosed the model size or parameter count".

Two things stand out. The Omniscience set measures knowledge questions: Ember answers 21.1% correctly, the lowest of the three, and its low hallucination rate comes mostly from declining to answer, about 65 questions in 100 (derived from accuracy and net score, a formula that reproduces Artificial Analysis's printed hallucination rates for all three models). For routing, extraction and repeated tool calls that may not matter; for lookups it does, and it matches Meta's own advice above about knowledge heavy questions.

The second is that the two independent cost figures disagree about the size of the gap. Artificial Analysis puts Ember at about a third of Haiku 5.5's cost per task. Vals AI's page puts Ember's cost per test at $0.15 against Haiku 5.5's $3.17, 21 times lower (derived), with Ember 1.75 points ahead on accuracy. The pages give different task mixes, token counts (23.13 million tokens per test for Ember against 38.15 million for Haiku 5.5) and no reasoning setting in the rows read. This site does not reconcile them. It reports both.

TechCrunch's report, dated 9 November 2026 by its page data, says Ember is "cost-competitive with open source competitors like DeepSeek V4.1 Flash and GLM 5.3 Flash". The measured pages say: Artificial Analysis lists DeepSeek V4.1 Flash 1017 (Max) at 32.59 and $0.6021 per task against Ember's 36.02 and $0.1359; Vals lists DeepSeek V4.1 Flash 1017 at 61.26% and $0.66 per test, 0.78 points above Ember's 60.48% at $0.15 (derived), and GLM 5.3 Flash at 50.86% and $2.05.

Small differences of labelling remain. Artificial Analysis's page calls its index v5.0.1 where Meta's report says v5.0 and its footnote v5; the page lists Ember's inputs as text and image where Meta lists text, image, audio and video; and its data marks Ember's price and performance source as "median" where Haiku 5.5's and Luna's are marked "firstParty", which the page does not explain beyond saying the price is "based on the median across providers serving the model" and that Ember has 1 provider.

What Meta's own pricing pages add to "about a third"

The headline price, $0.10 and $0.50 per million tokens with a 1M-token context window, has a condition on Meta's pricing page: "A request with up to 200K input tokens is billed at the rates in the first column. A request with more is billed at the long-context rates in full, not only for the tokens past 200K". Those rates are $0.30 input and $1.50 output, 3 times the headline (derived), for the whole request. Meta's launch post says Ember can take in a full codebase in one call; a request that does so crosses the line. Meta's choice page says to "compare costs at your real input sizes".

Other terms on the same pages: when a request sets no reasoning effort, Ember runs at max, the most expensive setting on Meta's chart (Meta's choose a model page); web search grounding adds "$2.50 per 1,000 search queries"; Meta "injects a small amount of steering context into every prompt" that is not billed; and the limits for muse-ember-1.0 are 3,000 requests and 4,000,000 tokens per minute, per team. Artificial Analysis's cost breakdown for Ember has a cache write line worth 19.6% of its cost per task (derived), yet Meta's pricing page prints no cache write rate; the OpenRouter listing prints $0.125 per million (raw 0.000000125 per token). Cache reads are $0.004 per million for Ember against $0.01 for Haiku 5.5 and Luna on Artificial Analysis's page, 2.5 times lower (derived).

On data use, Meta's pricing page says that in Standard mode "your content is not used to train Meta models" and that "Contributor mode isn't available for Muse Wonder or Muse Ember on Model API". Ember runs in Standard mode only. The same page says Muse Code subscriptions default to Contributor mode, switchable in settings; the text read does not say how that setting applies to Ember requests, so that is a question for the terms, not for this brief. (Any reading that Ember's price matches Meta's discounted Contributor tier is not supported by these pages: Contributor pricing is listed only for Muse Spark.)

What Meta states, and what it does not

On safety, Meta's post says Ember was evaluated "under our Meta Superintelligence Scaling Framework" and "remained below the framework’s high-risk thresholds" in the categories assessed. The framework itself is described in a 2 October 2026 Meta post. The only document the launch page links is the four page benchmark methodology; no safety report or system card for Ember was found in the pages read.

Five companies are quoted. Box praises "impressive accuracy, speed and token efficiency" on its own eval; Cloudflare says Ember "up to 30% faster at the tail than the comparable model we tested"; You.com reports a p95 latency of 2.2 seconds, "with the lowest latency of any model we tested". None prints a cost per task. Cloudflare's wording compares against an unnamed model.

Meta's developer blog shows its own spend on two worked examples: a 16 lexer code migration in which Ember handled 158 million input tokens at $1.84 of a $3.96 run, and a photo job whose Ember cost is printed as $0.50 in one sentence and $0.52 in another, with totals of $1.59 and $1.33. These are Meta's numbers at pay as you go prices and the post adds "Results vary from run to run".

Labels doing the work of evidence

"Intelligence per dollar." The post gives no formula. Dividing index by cost (this site's reading, derived) gives 293.2 for Luna's cheapest setting, 292.2 for Ember's xhigh and 263.2 for Ember's max, so by that ratio Luna's cheapest point edges Ember. The index is not a ratio scale, so this site does not lean on it; the point is that the phrase names a measure the page never defines.

"Smallest." Meta calls Ember its smallest model, but prints no size. The label says nothing on accuracy: Artificial Analysis shows 21.1% on knowledge questions, the lowest of the three.

"Estimates at proposed pricing." The caption marks the cost column as Meta's own, but the per token price behind it is the same $0.10 and $0.50 that Meta's pricing page bills. What stays an estimate for a buyer is the token count on a benchmark mix that is not your workload.

"1M-token context window." True of the model, and a 3 times price step above 200K input tokens.

Where the sources differ

What to do, in order (this site's suggestions)

Technical details

Meta's launch post prints no code and no configuration. The code blocks below hold what Meta's developer pages, the OpenRouter listing and Artificial Analysis's page data print, copied as read on 9 November 2026. No working exploit or credential is involved.

Prices and limits. This is Meta's pricing for the one model id, with the long context tier. A defender or buyer uses it to price a workload at its real request sizes, not the headline.

Muse Ember: muse-ember-1.0, in Standard mode only.
Usage  Up to 200K input tokens (per 1M tokens)  Over 200K input tokens (per 1M tokens)
Cached input  $0.004  $0.012
Input  $0.10  $0.30
Output  $0.50  $1.50
Rate limits: muse-ember-1.0  3,000 RPM  4,000,000 TPM

The public model listing. OpenRouter's public models listing for meta/muse-ember prints these fields. Note the maximum completion of 943,718 tokens against Meta's 524,288, and the null Hugging Face id. Use it to check a router's configuration against Meta's.

"id": "meta/muse-ember"
"context_length": 1048576
"modality": "text+image+file+audio+video->text"
"prompt": "0.0000001"
"completion": "0.0000005"
"web_search": "0.0025"
"input_cache_read": "0.000000004"
"input_cache_write": "0.000000125"
"max_completion_tokens": 943718
"is_moderated": true
"hugging_face_id": null
"default_effort": "max"
"mandatory": true

Artificial Analysis page data. These key and value pairs are copied from the data embedded in the Artificial Analysis models page for the three max effort entries, in the order Ember, Haiku 5.5, Luna. They are the source of the independent table above.

"name": "Muse Ember (Max)"
"intelligenceIndex": 36.0212987864912
"omniscienceAccuracy": 0.21066666666666667
"omniscienceHallucinationRate": 0.1735641891891892
"terminalBench40": 0.409090909090909
"frontierCode20": 0.271696212121212
"price1mInputTokens": 0.1
"price1mOutputTokens": 0.5
"cacheHitPrice": 0.004

"name": "Claude Haiku 5.5 (Max)"
"intelligenceIndex": 36.4845435880142
"omniscienceAccuracy": 0.3466666666666667
"omniscienceHallucinationRate": 0.3423469387755102
"terminalBench40": 0.328282828282828
"frontierCode20": 0.349325505050505
"price1mInputTokens": 0.1
"price1mOutputTokens": 0.5
"cacheHitPrice": 0.01

"name": "GPT-6 Luna (Max)"
"intelligenceIndex": 27.8552004938153
"omniscienceAccuracy": 0.3963333333333333
"omniscienceHallucinationRate": 0.7934842628382109
"terminalBench40": 0.126262626262626
"frontierCode20": 0.251747727272727
"price1mInputTokens": 0.1
"price1mOutputTokens": 0.5
"cacheHitPrice": 0.01

Meta's chart values. The nine points as the launch page prints them, in the page's order. Setting labels are printed separately, so the pairing in the article is this site's.

Muse EmberCost Per Task (USD)$0.116
Muse EmberScore33.9
Muse EmberCost Per Task (USD)$0.136
Muse EmberScore35.8
Claude Haiku 5.5Cost Per Task (USD)$0.092
Claude Haiku 5.5Score24.6
Claude Haiku 5.5Cost Per Task (USD)$0.159
Claude Haiku 5.5Score29.6
Claude Haiku 5.5Cost Per Task (USD)$0.248
Claude Haiku 5.5Score34.3
Claude Haiku 5.5Cost Per Task (USD)$0.411
Claude Haiku 5.5Score36.2
GPT-6 LunaCost Per Task (USD)$0.073
GPT-6 LunaScore21.4
GPT-6 LunaCost Per Task (USD)$0.104
GPT-6 LunaScore23.3
GPT-6 LunaCost Per Task (USD)$0.152
GPT-6 LunaScore27.8

Meta's evaluation table, with qualifiers. Printed as the page prints it, including "best result" markers and the partial and binary OSWorld scores. Treat the Terminal-Bench 4.0 row and the three internal rows with the cautions above.

AutomationBench-AA   E2E business workflows   Best result. 37.8  13.9  20.2
Agents' Last Exam   End-to-end agent tasks   Best result. 61.6  57.0  57.1
Personal Superintelligence Index   Internal composite index of personal and knowledge work   Best result. 46.8  36.1  37.0
Terminal-Bench 4.0   Agentic terminal coding, best published harness   Best result. 40.9  39.2  16.4
Coding Agent Index-AA   Coding in real agent harnesses   Best result. 46  37  31
FrontierCode 2.0-AA   Real engineering tasks judged on mergeability   27.2  Best result. 34.9  25.2
GDP.pdf-AA   Professional work over long documents   Best result. 22.8  20.8  Best result. 22.8
OSWorld 2.1   Computer use across desktop apps   Best result. 69.5 (partial) 33.8 (binary)  68.9 (partial) 27.8 (binary)  50.5 (partial) 17.0 (binary)
PerceptionBench (w/ tools)   Fine-grained visual perception   Best result. 80.6  78.6  62.8
Artificial Analysis Index   Best result. 36  Best result. 36  28

Meta's footnote. The qualifiers that decide how to read the table: harnesses, composite, partial credit, tools.

1.
All models run at max reasoning effort.
Artificial Analysis results are from Intelligence Index v5, as published November 9, 2026; cost per task uses each model's list pricing.
Internal results are from Meta's evaluation suite.
The Personal Superintelligence Index (PSI) is an internal composite of eight evaluations: Agents' Last Exam, AutomationBench, FocusBench, JobBench, OfficeQA Pro, GDP.pdf, GDP.xlsx and VBT.
Terminal-Bench 4.0 uses each model's best published harness: mini-swe-agent (Muse Ember), Claude Code (Claude Haiku 5.5), and Codex (GPT-6 Luna).
OSWorld 2.1 uses partial-credit reward; PerceptionBench is scored with tools.

Meta's first API call for Ember. From Meta's developer documentation. A buyer can reuse it to test the model with a key held in an environment variable; the key is never written into the command.

curl -X POST "https://api.meta.ai/v1/responses" \
  -H "Authorization: Bearer $MODEL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "model": "muse-ember-1.0", "input": "Classify this support ticket as billing, bug, or feature request, and give a one-sentence reason: \"I was charged twice for my October invoice.\"" }'

Meta's example worker, and the flag to avoid. The photo sorting example runs agents on Muse Ember under a larger model. Meta's text says the flag below approves shell commands without asking and lets the agents read the API key. A defender treats this as a demo, not a deployment pattern.

"ember-worker": {
  "description": "A muse-ember worker for one bounded, fully specified task. Give it the complete task: the files it owns, the commands to run, and what to report.",
  "model": "muse-ember",
  "effort": "low",
  "maxTurns": 20,
  "prompt": "Do exactly the task in your input. Edit only the files it names, run the commands it gives, fix your own files if a command fails, and report the result in one short line. Never claim a result you did not run."
}

muse --model muse-wonder-1.0 --reasoning-effort low \
  --agents "$AGENTS" -c agents.execution_capacity=17 \
  -c 'run.workflow_trigger_mode="off"' \
  --parallel-tool-calls --yolo \
  "$PROMPT"

No hashes, indicators, detection rules or log lines are printed by any of these sources, and none are invented here.

The question Meta leaves open

Meta prints a price per million tokens and a cost per task, and leaves out the number that connects them: the tokens Ember spends per task. If the whole claim is that Ember uses a third of the tokens at the same price, how many does it spend on your tasks, at which reasoning setting, and under how many tokens of prompt?

Key facts