Meta's Muse Ember Scores 36 at $0.14 per Task at Max Reasoning on Artificial Analysis, Which Reports 68.1% on AutomationBench-AA Versus Meta's 37.8%

Overview

Meta introduced Muse Ember on November 9, 2026. Its announcement calls it “our smallest and most efficient model,” built for the repeated steps inside agents: tool calls, coding subtasks, routing, extraction and long document parsing. Meta’s Model API documentation lists it as muse-ember-1.0. Artificial Analysis, the evaluator whose results Meta’s comparison charts use, scores Muse Ember (max reasoning) 36 on its Intelligence Index at $0.14 per task. Meta’s and Artificial Analysis’s published figures match on that score and cost but differ on two benchmark rows, covered below.

What Meta Says

API and availability. According to the model page, Muse Ember is available on the Meta Model API and in Muse Code. It accepts text, image, video, audio and PDF input and returns text, with a context window of 1,048,576 tokens and a maximum output of 524,288 tokens. Reasoning effort runs from “minimal” to “max,” and the default is “max.” On the Model API it runs in Standard mode only; the page says “it has no Contributor-mode variant.” Meta’s pricing page describes Contributor mode as a discounted tier offered in exchange for permission to use a customer’s content to train future Meta models.

Pricing, as read on November 10, 2026. Meta’s pricing page lists, per million tokens, $0.10 input, $0.004 cached input and $0.50 output for requests of up to 200K input tokens. Above 200K the rates are $0.30, $0.012 and $1.50, and a larger request “is billed at the long-context rates in full, not only for the tokens past 200K.” Web search grounding adds $2.50 per 1,000 search queries on top of token costs, and the page sets muse-ember-1.0 limits of 3,000 requests and 4,000,000 tokens per minute per team. On the same page, Muse Spark’s Standard-mode rates are $1.25 input and $4.25 output, and Muse Wonder’s are $4.00 and $20.00; the Herald previously reported Muse Wonder’s API launch.

Positioning. Meta’s announcement says Muse Ember is the first model it trained using Muse Wonder Full, and places it as Muse Wonder for the hardest problems, Muse Spark for coding and agentic workflows, and Muse Ember for repeatable tasks at scale. The documentation tells developers to use Muse Spark or Muse Wonder for hard single-shot coding problems, knowledge-heavy questions and visual coding. The announcement also says Meta’s pre-release safety evaluations, run under its Meta Superintelligence Scaling Framework, found Muse Ember “remained below the framework’s high-risk thresholds” across the categories assessed.

The Benchmark Figures

Meta’s announcement compares Muse Ember with Claude Haiku 5.5 and GPT-6 Luna. Its table uses max reasoning effort, and Meta says Muse Ember matches Haiku’s best score at about a third of the cost and beats every Luna setting. In Meta’s chart, Muse Ember (max) scores 35.8 at $0.136 per task, against 36.2 at $0.411 for Haiku (max); Muse Ember (xhigh) scores 33.9 at $0.116.

Artificial Analysis reports the same pairing in its own words: Muse Ember (max) is “the cheapest model scoring 36 or higher,” and Haiku (max) costs almost 3x more at $0.41 for the same score. It adds that Muse Ember uses roughly a third as many output tokens as Haiku and that its cached input is 60% cheaper. The Decoder reports that Muse Spark 1.3 costs $2.39 per task for a nearly identical score, 37 against Muse Ember’s 36, and that Haiku outperforms Muse Ember on FrontierCode 2.0-AA, 34.9 percent to 27.2 percent.

Rows from Meta’s table, as Meta reports them (Muse Ember, Haiku, Luna): Terminal-Bench 4.0 at 40.9, 39.2 and 16.4; Coding Agent Index-AA at 46, 37 and 31; FrontierCode 2.0-AA at 27.2, 34.9 and 25.2. Meta labels its OSWorld 2.1 row (partial credit: 69.5, 68.9 and 50.5) and its Agents’ Last Exam row (61.6, 57.0 and 57.1) as Meta internal evaluations, and describes its Personal Superintelligence Index row (46.8, 36.1 and 37.0) as an internal composite.

Artificial Analysis’s article also reports a weak spot that the announcement’s results table does not list: Muse Ember’s AA-Omniscience accuracy is 21.1%, the lowest of the similarly intelligent models it compared, against 34.7% for Haiku (max). It says this is mostly because the model declines to answer 64% of questions, which keeps its hallucination rate at 17.4%, half of Haiku’s. On its model page, read November 10, Artificial Analysis lists an output speed of 448 tokens per second and 3.1 minutes per Intelligence Index task.

Where the Sources Differ

• AutomationBench-AA. Meta’s table gives Muse Ember 37.8%, with Luna at 20.2% and Haiku at 13.9%, marked “Strict pass@1.” Meta’s evaluation report describes the row as Artificial Analysis’s run of Zapier’s private set of 657 tasks, scored as the share completed in full. Artificial Analysis’s article says Muse Ember scores 68.1% on AutomationBench-AA, “level with Muse Wonder (max)”. Its article does not state the metric behind that figure, and neither source’s text explains the gap. The data embedded in the page source of Artificial Analysis’s model page stores Muse Ember’s AutomationBench-AA value (0.6808) under a field named automationBenchPartialScore, which points to a partial-credit measure rather than the share of tasks completed in full; the Herald read this from the page source, and Artificial Analysis does not say so in its article.
• Terminal-Bench 4.0. Both sources give Muse Ember 40.9%. For Haiku, Meta gives 39.2% and its report says this is Anthropic’s self-reported Claude Code result from the Claude Haiku 5.5 system card, while Muse Ember’s figure is Artificial Analysis’s run in the mini-swe-agent harness. Artificial Analysis gives Haiku 32.8%, in a passage about its Intelligence Index evaluations, without naming the harness. Meta’s report describes a self-reported vendor result for its Haiku figure; Artificial Analysis’s article does not say how its figure was obtained.
• Input types. Meta’s documentation lists text, image, video, audio and PDF input. The Decoder lists text, images, audio and video. Artificial Analysis’s model page says the model “supports text and image input”. None of the cited pages reconciles these.
• Cost basis. Meta’s chart caption says “Muse Ember costs are estimates at proposed pricing,” while its footnote says cost per task “uses each model’s list pricing.” Meta’s footnote cites Intelligence Index v5.0; Artificial Analysis’s model page now shows v5.0.1.

Weights

Meta’s models documentation lists Muse Ember among its API-hosted models and names Muse Glimmer and Muse Spark Open as its open-weight lines. The New Stack reports that Meta “is keeping Ember’s model weights private,” and Artificial Analysis’s model page labels it a “Proprietary model.” The Meta pages read for this article give no parameter count.

What We Don’t Know

• Whether Artificial Analysis’s 68.1% on AutomationBench-AA is the partial-credit measure its page data suggests, since its article does not name the metric.
• Whether Artificial Analysis’s Haiku run on Terminal-Bench 4.0 used a different harness from the Claude Code result Meta cites.
• Whether the cost-per-task gaps hold for workloads outside Artificial Analysis’s task set; Meta’s footnote says its cost per task uses list pricing, and Meta’s report says its internally generated competitor runs “remain best-effort.”