> ## Content Index
> Fetch the complete content index at: https://www.tao.media/llms.txt
> Use this file to discover other available public pages before exploring further.

# Meta Launches Muse Ember for High-Volume AI Agents at One-Third of Haiku’s Task Cost
- URL: https://www.tao.media/meta-launches-muse-ember-for-high-volume-ai-agents-at-one-third-of-haikus-task-cost/
- Published: 2026-11-09T20:13:28.000Z
- Updated: 2026-11-09T21:29:11.000Z
- Description: The new worker model combines million-token context with lower token use to cut the cost of repeated coding and agent tasks.
- Author: Bart Hillerich
- Tags: Meta, AI, News

Meta has released Muse Ember, a low-cost AI model built for the repeated tool calls, coding subtasks, and document processing that make up much of an agent's workload. It matches Claude Haiku 5.5's rounded score on the Artificial Analysis Intelligence Index while costing about one-third as much per benchmark task.

[Mark Zuckerberg announced the release](https://x.com/finkd/status/2119851180948152590?ref=tao.media), calling Ember Meta's smallest model. It is available through the Meta Model API and Muse Code, giving developers a cheaper option for work that would otherwise consume repeated calls to a larger model.

Meta says Ember is the first model trained using Muse Wonder Full, extending the development strategy introduced with the Muse Wonder model family. The company is now dividing its lineup between frontier reasoning, coding and agent workflows, and high-volume execution.

[Meta launches Muse Wonder, its first frontier model family | All Things MuseMark Zuckerberg announced Muse Wonder and Muse Wonder Full on X: Meta's first frontier-class models, available via Muse subscription, Muse Code, and the Meta![](https://storage.ghost.io/c/78/0b/780ba906-b1a7-4bf0-873c-bdd5c32e5331/content/images/icon/favicon-48024383-c21e-43cf-9528-94fce4f298aa.png)All Things MuseIcarus![](https://storage.ghost.io/c/78/0b/780ba906-b1a7-4bf0-873c-bdd5c32e5331/content/images/thumbnail/cover-meta-muse-wonder-frontier-model-5642f654-a230-43c4-92fb-e831418ed482.webp)](https://www.allthingsmuse.com/articles/meta-muse-wonder-frontier-model/?ref=tao.media)

## Lower Task Costs Without Lower Token Prices

The central result is independently supported by [Artificial Analysis' evaluation](https://x.com/ArtificialAnlys/status/2119854348587123007?ref=tao.media). At maximum reasoning effort, Ember scores 36 on the firm's Intelligence Index and costs about $0.14 per task, compared with Claude Haiku 5.5's score of 36 at about $0.41\. Artificial Analysis identifies Ember as the cheapest model scoring 36 or higher in its comparison.

The more precise figures in [Meta's cost comparison](https://x.com/AIatMeta/status/2119853595453702580?ref=tao.media) are 35.8 points at $0.136 per task for Ember and 36.2 points at $0.411 for Haiku. GPT-6 Luna's best setting scores 27.8 at $0.152 per task.

![Intelligence vs. Cost Per Task chart](https://storage.ghost.io/c/78/0b/780ba906-b1a7-4bf0-873c-bdd5c32e5331/content/images/2026/11/HWs6A9VbIAEjodf.jpg)

[Intelligence vs. Cost Per Task chart | Meta (@AIatMeta on X)](https://x.com/AIatMeta/status/2119853595453702580?ref=tao.media)

Ember has the same input and output token prices as Haiku 5.5 and GPT-6 Luna. Its lower bill comes from using fewer tokens to complete the evaluated tasks, instead of charging less for each token. Shorter execution can reduce spending across many steps for an agent that repeatedly reads files, calls tools, and generates intermediate responses.

The figures measure cost on the benchmark's workload, not a fixed fee for any developer task. Meta's [API pricing](https://x.com/AIatMeta/status/2119853597542568405?ref=tao.media) is $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens.

Within Meta's lineup, the gap is larger:

| Model at Maximum Reasoning | Artificial Analysis Intelligence Index | Cost Per Index Task |
| -------------------------- | -------------------------------------- | ------------------- |
| Muse Ember                 | 36                                     | $0.14               |
| Muse Spark 1.3             | 37                                     | $2.39               |
| Muse Wonder                | 44                                     | $6.95               |

Meta Chief AI Officer Alexandr Wang [compared Ember with Muse Spark 1.3](https://x.com/alexandr%5Fwang/status/2119857909098283406?ref=tao.media), pointing to nearly the same index score at roughly one-seventeenth of the cost per task. Muse Wonder retains an eight-point lead over Ember, at about 50 times the benchmark task cost.

![Muse Ember vs. Muse Spark 1.3 on Artificial Analysis](https://storage.ghost.io/c/78/0b/780ba906-b1a7-4bf0-873c-bdd5c32e5331/content/images/2026/11/HWs_iJba0AEZwRa.jpg)

[Muse Ember vs. Muse Spark 1.3 on Artificial Analysis | Alexandr Wang (X)](https://x.com/alexandr%5Fwang/status/2119857909098283406?ref=tao.media)

## A Worker Model With Million-Token Context

[Muse Ember](https://research.meta.ai/blog/muse-ember?ref=tao.media) supports a 1 million-token context window and up to 512,000 output tokens per response. The context window is the amount of material a model can consider in a call, including instructions, documents, and conversation history; the separate output limit governs how much it can generate.

Its native inputs include text, images, audio, and video. That lets an agent work from screenshots, scanned documents, charts, voice recordings, and video without sending each input type to a different model. Meta is targeting extraction, routing, long-document parsing, and coding work such as reading files, running tests, and making targeted fixes.

Wang [suggests pairing the models](https://x.com/alexandr%5Fwang/status/2119853377932956070?ref=tao.media) so that Muse Wonder handles planning while Ember takes on the volume of execution. In that arrangement, a larger model can set the approach and delegate narrower steps to the cheaper worker.

[Wang's launch thread](https://x.com/alexandr%5Fwang/status/2119853374158032950?ref=tao.media) places Ember at the small end of the Muse family. Muse Spark remains the coding and agent-workflow tier; Meta also released [Muse Spark Open with downloadable weights](https://tao.media/meta-opens-muse-spark-weights-for-coding-agents-and-two-mac-local-inference/?ref=tao.media) in October. Ember's release is through Meta's hosted API and Muse Code.

## Coding Strengths and a Knowledge Tradeoff

The benchmark results vary by task. Ember scores 46 on the Artificial Analysis Coding Agent Index when running in Muse Code, ahead of Haiku 5.5 in Claude Code at 37 and GPT-6 Luna in Codex at 31.

On Terminal-Bench 4.0, which tests work in a terminal environment, Ember scores 40.9%. Meta compares that with Haiku's 39.2% and Luna's 16.4%, using each model's best published harness, the software that supplies tools and manages the agent loop. Those runs use mini-swe-agent for Ember, Claude Code for Haiku, and Codex for Luna.

Artificial Analysis reports a different Haiku result, 32.8%, in its own Intelligence Index testing. That is a separate comparison from Meta's best-published-harness table. On FrontierCode 2.0, which judges real engineering changes on whether they are suitable to merge, both sources put Ember at 27.2%, behind Haiku's 34.9% and ahead of Luna's 25.2%.

![Muse Ember evaluation results table](https://storage.ghost.io/c/78/0b/780ba906-b1a7-4bf0-873c-bdd5c32e5331/content/images/2026/11/HWs55-3bYAEaslZ.png)

[Muse Ember evaluation results table | Meta (@AIatMeta on X)](https://x.com/AIatMeta/status/2119853591703998787?ref=tao.media)

For document work, Ember scores 22.8% on GDP.pdf, matching Luna and exceeding Haiku's 20.8%. The [evaluation methodology](https://research.meta.ai/static/muse-ember-evaluation-report?ref=tao.media) explains that Artificial Analysis supplies extracted PDF text and page images, with success requiring an answer to meet every grading criterion.

Meta's internal evaluations cover additional capabilities. The company reports 61.6% on the Linux subset of Agents' Last Exam, 69.5% partial credit on OSWorld 2.1 computer-use workflows, and 80.6% on PerceptionBench with tools. Those measure different forms of task completion and visual understanding, rather than a single interchangeable accuracy score.

General knowledge is a weaker area. Artificial Analysis measured 21.1% accuracy on AA-Omniscience, compared with Haiku's 34.7%. Ember declined to answer 64% of questions, contributing to a 17.4% hallucination rate, about half Haiku's. Its lower knowledge accuracy therefore comes with a greater tendency to leave questions unanswered rather than supply an incorrect answer.

## Partner Tests and Developer Access

Early partners have tested Ember on enterprise content and coding. Box's vice president of AI products, Yashodha Bhavnani, said it could assess contracts clause by clause against negotiation playbooks and reach the appropriate overall risk rating. Cline founder Saoud Rizwan highlighted coding and document-analysis work delegated to subagents.

Cloudflare CTO Dane Knecht reported up to 30% faster performance at the slower end of its response times than the comparable model it tested. You.com's inference lead, Chak Pothina, reported a 95th-percentile latency of 2.2 seconds on coding tests, with 36% to 48% fewer output tokens. Those are partner-reported results from their respective tests.

Meta says Ember stayed below the high-risk thresholds in its Meta Superintelligence Scaling Framework assessments of chemical and biological, cybersecurity, and loss-of-control risks. It also tested whether the model follows user intent, protects sensitive information, resists instructions hidden in external content, and seeks human input before high-impact actions.

Developers can use Ember now under the model name `muse-ember-1.0`, with support for the Responses API. Meta says more efficient models are in training.