Meta Releases Muse Ember, Pitching Lower Cost per Task for Agent Work
Meta prices its new agent model at $0.10 input and $0.50 output per million tokens on requests up to 200K input tokens; its efficiency edge over Claude Haiku 5.5 and GPT-6 Luna is a per-task claim from Meta's own charts, not a lower per-token price.
Published
Last updated / Fact-checked on
Illustrative photo of glowing charcoal in a grill; it is not Meta, Muse Ember, or any compute hardware. Photo: ClickerHappy / Pixabay, Pixabay Content License.
Meta on Monday, November 9, released Muse Ember (muse-ember-1.0), which it calls its smallest and most efficient model, built for "the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing." Meta says it is available now on the Meta Model API and in Muse Code, at $0.10 per million input tokens and $0.50 per million output tokens for requests up to 200K input tokens. Longer requests are billed at three times those rates.
Meta's headline claim is that Ember delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna. That claim is about cost per task, drawn from Meta's own charts, with Ember's costs labeled by Meta as estimates. The New Stack reports that Ember's per-token rates match both rivals' standard short-context prices.
This follows DailyMech's coverage of specialized consumer models trained using Muse Wonder and Spaces in the Muse app. What is new: Ember is a named developer model, and Meta says it is the first trained using Muse Wonder Full.
At a glance
| Item | Detail |
|---|---|
| Model ID | muse-ember-1.0 |
| Where | Meta Model API, Muse Code |
| Up to 200K input | $0.10 input / $0.004 cached / $0.50 output per 1M tokens |
| Over 200K input | Whole request: $0.30 / $0.012 / $1.50 per 1M |
| Web search grounding | $2.50 per 1,000 queries, extra |
| Context / output | 1,048,576 / 524,288 tokens |
| Inputs | Text, image, video, audio, PDF; text out |
| Lab | Meta Superintelligence Labs |
What $0.10 actually buys
The 10-cent figure is the short-context input rate only. Meta's pricing page bills any request over 200K input tokens entirely at the higher tier, not just the tokens past the threshold. Output costs five times input in both tiers, and web search grounding is billed separately.
Reasoning effort runs from "minimal" to "max," and "max" is the default. Meta lists rate limits of 3,000 requests and 4 million tokens per minute per team. Ember runs in Standard mode only: Meta's docs say the Contributor mode, which trains on customer data, is not available for Ember on the Model API.
Meta's comparison, and its limits
| Claim | Source | Status |
|---|---|---|
| More intelligence per dollar than Haiku 5.5, Luna | Meta's charts | Meta's claim; Ember costs estimated |
| Haiku's best index score at about a third of the cost | Meta's charts, Artificial Analysis data | Not independently verified |
| Same per-token rates as Haiku 5.5, Luna | The New Stack | Rival price pages not checked |
Meta reports that Ember scores 36 on the Artificial Analysis Intelligence Index, matching Haiku's best, at about $0.14 per task. The New Stack reads Meta's chart as roughly $0.41 for Haiku. These are Artificial Analysis figures as Meta chose to show them; Artificial Analysis' own Ember page returned a 404 when DailyMech checked.
Meta's materials carry their own caveats. One chart note calls Ember's costs estimates at "proposed pricing," another at "list pricing." Its evaluation report says competitor runs were best-effort and may not be optimally tuned, and Terminal-Bench used different harnesses per model.
Not every Meta-reported result favors Ember. On FrontierCode 2.0, Meta reports Ember at 27.2 versus Haiku's 34.9. On AutomationBench-AA, it reports Ember at 37.8, Haiku at 13.9 and Luna at 20.2.
So Ember is not shown to be cheaper per token than Haiku or Luna. Meta's argument is that Ember costs less per completed task.
Where Ember fits in the Muse lineup
Meta's shorthand is "Muse Wonder plans, Muse Ember handles the volume," and its docs suggest pairing Ember with Muse Spark or Muse Wonder. Against its own Muse Spark 1.3, Meta reports a score of 37 versus Ember's 36, at $2.39 versus $0.14 per task.
Meta says Ember stayed below the high-risk thresholds of its Superintelligence Scaling Framework for chemical and biological, cyber and loss-of-control risks. That is Meta's own assessment.
Meta has not disclosed Ember's parameter count, architecture, training method, regional availability, consumer-app use or third-party distribution. The New Stack says the weights are private.
What to watch
Whether Artificial Analysis publishes its own Ember listing; independent tests and latency measurements; availability on other platforms; any open-weight release or separate safety report; and pricing responses from Anthropic or OpenAI.
How we reported this
On November 9 we read Meta's announcement, Model API docs, pricing page and evaluation report, plus TechCrunch and The New Stack. We re-read our November 6 and November 8 Muse articles for consistency. Artificial Analysis' Ember page returned a 404. We excluded CryptoBriefing because its account conflicted with Meta's primary pages. We did not test the model, check Anthropic's or OpenAI's price pages, or contact anyone.
Published and fact-checked November 9, 2026.
Sources
- Meta AI Research, "Introducing Our Smallest, Most Efficient Model: Muse Ember," November 9, 2026 (date only): https://research.meta.ai/blog/muse-ember (fully readable; charts rendered as tables)
- Meta Model API docs, "Muse Ember": https://dev.meta.ai/docs/muse-ember (fully readable)
- Meta Model API, "Pricing and rate limits": https://dev.meta.ai/docs/pricing-rate-limits (fully readable)
- Meta, "Muse Ember Evaluation Report": https://research.meta.ai/static/muse-ember-evaluation-report (fully readable; methodology)
- TechCrunch, Russell Brandom, "Meta releases Muse Ember, a new lightweight model for agent execution," November 9, 2026, 10:00 AM PST: https://techcrunch.com/2026/11/09/meta-releases-muse-ember-a-new-lightweight-model-for-agent-execution/ (fully readable; short)
- The New Stack, Frederic Lardinois, "Meta launches Muse Ember to take on Claude Haiku and GPT-6 Luna," November 9, 2026 (no time zone shown): https://thenewstack.io/meta-muse-ember-launch/ (fully readable)
- Artificial Analysis, Muse Ember model page: https://artificialanalysis.ai/models/muse-ember (returned 404; not read)
- CryptoBriefing, Muse Ember launch article: https://cryptobriefing.com/meta-launches-muse-ember-reasoning-model/ (read; excluded as contradicted by Meta's primary pages)
DailyMech used AI tools to help research and draft this article; it was fact-checked by our editors.