Meta released Muse Ember 1.0 on November 9, 2026, positioning it as the smallest, most efficient model in its Muse lineup. Its assigned work is specific: tool calls, coding subtasks, routing, extraction and long-document parsing, the repeated steps that can make an agent expensive to operate at scale.
The model is available through the Meta Model API and Muse Code. Meta advertises a 1 million-token context window, multimodal input and starting prices of $0.10 per million input tokens and $0.50 per million output tokens. Those prices come with an important condition: requests above 200,000 input tokens move into a tier costing three times as much. Meta’s announcement supplies the headline specifications; VentureBeat’s launch coverage details the pricing threshold.
The useful distinction is between cheap tokens and cheap completed tasks. Ember’s base input and output rates match Claude Haiku 5.5 and GPT-6 Luna. Its stronger economic argument comes from published evaluations showing lower spending per task through cheaper cached input and fewer generated tokens.
Where Ember Fits in an Agent System
Meta designed Ember for execution inside a larger agent architecture. A more capable planner can establish a goal, break it into steps and delegate bounded jobs to a cheaper worker.
TechCrunch describes a customer-support example: a larger model establishes handling rules, then a smaller model applies them across tens of thousands of cases. Meta positions Muse Wonder as the planner and Ember as the high-volume execution layer.
A hypothetical implementation could divide responsibilities this way:
- Planner: Interpret an unusual support problem and decide which policy applies.
- Ember worker: Retrieve records, extract relevant details and prepare a structured response.
- Validation layer: Check required fields and whether the proposed action is permitted.
- Escalation path: Return ambiguous or consequential decisions to a stronger model or a person.
That division is an architectural choice, not a requirement to use Ember. It can also power a standalone workflow.
I’d start with tasks whose success can be checked cheaply, such as extracting contract fields or making a small code change that must pass tests. Low inference prices are most useful when mistakes are detectable. An inexpensive worker that silently updates the wrong record can still produce an expensive outcome.
What Developers Can Use Today
Meta’s announcement confirms availability on its API and Muse Code. VentureBeat reports the following implementation details from the launch documentation:
| Feature | Published specification |
|---|---|
| API model identifier | muse-ember-1.0 |
| API interfaces | Responses, Chat Completions and Messages |
| Input types | Text, images, audio, video and PDFs |
| Output type | Text |
| Context window | 1,048,576 tokens, advertised as 1M |
| Maximum output | 524,288 tokens, advertised as 512K |
| Reasoning effort | Minimal through max, with max the default |
These details come from VentureBeat’s documentation review; Meta’s announcement directly confirms the headline context, output and multimodal capabilities.
Multimodal input gives the worker access to material that would otherwise require another processing step: screenshots, scanned documents, meeting recordings or video. Text output means developers should not read that capability as image, audio or video generation.
The large context window is a capacity limit, not a recommendation to fill every request. It can accommodate substantial document collections, but sending unnecessary material increases input spending and can push a request into the higher pricing tier.
Likewise, the maximum output allowance is a ceiling. An extraction task returning a few structured fields has very different economics from an agent generating a long analysis.
The launch describes hosted access and provides no open-weight download. Developers planning self-hosting should therefore treat this release as an API offering.
The Pricing Catch Is Above 200,000 Input Tokens
Ember’s low headline price applies to requests containing up to 200,000 input tokens. Above that threshold, uncached input, cached input and output rates all triple.
| Model and request tier | Input per 1M tokens | Output per 1M tokens |
|---|---|---|
| Muse Ember, up to 200K input | $0.10 | $0.50 |
| Muse Ember, above 200K input | $0.30 | $1.50 |
| Claude Haiku 5.5, up to 100K input | $0.10 | $0.50 |
| Claude Haiku 5.5, above 100K input | $0.50 | $2.50 |
| GPT-6 Luna, reported base rates | $0.10 | $0.50 |
These are the launch-period rates reported by VentureBeat. They are separate input and output charges, not a single blended price.
Ember’s cached-input rate is $0.004 per million tokens in the lower tier and $0.012 above 200,000 input tokens. Artificial Analysis reports that its base cached-input rate is 60% below those of Haiku 5.5 and Luna. That matters in agent workflows that repeatedly reuse cache-eligible context. Artificial Analysis explains the comparison here.
The Higher Rate Applies to the Whole Request
VentureBeat reports that crossing the threshold reprices the entire request, including its output. Cached input and conversation history carried through previous_response_id count toward the cutoff.
Using those published rates:
- 200,000 uncached input tokens plus 10,000 billed output tokens: $0.025.
- 200,001 uncached input tokens plus the same output: approximately $0.075.
Those are calculated model charges, excluding any separate application or tool costs. One additional input token triggers roughly a threefold bill in this example.
That creates a practical distinction between what fits and what remains in the cheapest tier. A million-token window does not make a million-token request eligible for the advertised starting rates.
I’d put a token-budget check before dispatching long requests. Removing irrelevant documents or retrieving only the sections needed for a task could avoid the price jump. Splitting work also has costs, including repeated instructions and additional coordination, so the useful metric is the total bill for a correctly completed workflow.
What the Benchmarks Actually Support
Meta’s launch chart says Ember sits on an efficiency frontier where “no model scores higher for less per task.” That is Meta’s presentation of benchmark results and estimated costs at its listed prices, not a universal guarantee across applications. The announcement identifies the chart as using Artificial Analysis results.
An independent launch evaluation is also available. Artificial Analysis reports:
| Model and reasoning setting | Intelligence Index, rounded | Cost per benchmark task |
|---|---|---|
| Muse Ember, max | 36 | About $0.14 |
| Muse Ember, xhigh | 34 | About $0.12 |
| Claude Haiku 5.5, max | 36 | About $0.41 |
| GPT-6 Luna, max | 28 |
Frequently Asked Questions
3 questions
1Is Muse Ember 1.0 Open-Weight?
The launch offers access through the Meta Model API and Muse Code. It does not describe an open-weight download.
