Meta releases Muse Ember for high-volume developer tasks
Developers can use the new model in Muse Code and through the Meta Model API, with Meta positioning it for repeated tasks alongside the larger Muse Wonder.
Loading… Բեռնվում է…
Developers can use the new model in Muse Code and through the Meta Model API, with Meta positioning it for repeated tasks alongside the larger Muse Wonder.
Meta released Muse Ember on 9 November 2026, giving developers a new superintelligence (SI) model for repeated jobs. It is available in Muse Code and through the Meta Model API as muse-ember-1.0, with support for the Responses API.
Ember follows Muse Spark in April, its open-source twin Muse Glimmer in August and Muse Wonder in October. Meta says Ember is the first model it trained using the larger Muse Wonder Full.
Meta says Ember is designed for tool calls, coding subtasks, routing, extracting information and parsing long documents. The company positions Wonder to plan work and Ember to carry out repeated tasks at volume.
Ember accepts text, images, audio and video. Meta says it can operate desktop applications using screenshots. Its context window—the material it can handle in a request—is 1 million tokens, and it supports up to 512K output tokens. Tokens are units used to measure model input and output.
| Token type | Price |
|---|---|
| Uncached input | $0.10 |
| Cached input | $0.004 |
| Output | $0.50 |
Cached input is previously processed material reused in a request. As The New Stack explains, Claude Haiku 5.5 and GPT-6 Luna match Ember’s uncached input and output rates for standard short-context requests. Both charge $0.01 per million cached input tokens at those rates and more for longer prompts.
Analysis
The company also announced Claude Mythos 5.2 on 9 November, with access initially limited to vetted organizations.
In Meta’s comparison on the Artificial Analysis Intelligence Index, a combined benchmark, Ember matched Haiku’s score at the max reasoning setting with a lower estimated cost per task. The task costs are comparison estimates, not developers' actual bills.
| Model | Score | Estimated cost per task |
|---|---|---|
| Muse Ember | 36 | About $0.14 |
| Claude Haiku 5.5 | 36 | $0.41 |
| GPT-6 Luna | 28 | $0.15 |
| Muse Spark 1.3 | 37 | $2.39 |
At Ember’s xhigh reasoning setting, Meta’s chart puts its score at about 34 and its cost at roughly $0.12 per task. The Spark estimate uses the standard API rates in the comparison.
Ember performed less strongly on FrontierCode 2.0, which assesses whether a software repository maintainer would merge an agent’s changes. Meta’s results put Ember below Haiku and DeepSeek, though above Spark.
| Model | Score |
|---|---|
| Muse Ember | 27.2% |
| Claude Haiku 5.5 | 34.9% |
| DeepSeek | 28.5% |
| Muse Spark 1.3 | 23.1% |
In Meta’s OSWorld 2.1 results for computer tasks, Ember received similar partial credit to Haiku and completed more tasks outright. Spark’s result came from a separate run.
| Model | Partial credit | Tasks completed |
|---|---|---|
| Muse Ember | 69.5% | 33.8% |
| Claude Haiku 5.5 | 68.9% | 27.8% |
| GPT-6 Luna | — | 17% |
| Muse Spark 1.3 | Nearly identical to Ember | 39.25% |
Meta is keeping Ember’s model weights—the numerical parameters that define the trained model—private. The New Stack expects Ember to reach services such as OpenRouter and neocloud providers, based on earlier Muse releases, but gives no date for that availability.
1 min
The Midnight Black model will cost ¥39,980 including tax and go on sale on 10 December. Sony says the storage is intended for future functions.
2 min
Sega said 92,273 records also included information associated with newsletter delivery, while SEGA Account addresses were outside the leak. It has suspended newsletters sent through provider Arara.
2 min