Meta Releases Muse Ember to Cut AI Agent Costs by 3x
Meta's new Muse Ember matches Claude Haiku 5.5 at a third the cost, with 1M context and native multimodal support for agent work.
- Meta launched Muse Ember, its smallest model, built for agent subtasks and tool calls.
- Priced at $0.10 per 1M input tokens, $0.50 per 1M output, $0.004 cached input.
- Matches Claude Haiku 5.5 on Intelligence Index at roughly one-third the cost per task.
- 1M-token context, 512K output, native text, image, audio, and video input.
- Leads on AutomationBench (37.8%), Terminal-Bench (40.9%), OSWorld 2.1 (69.5%); trails Haiku on FrontierCode 2.0.
- First model distilled from Muse Wonder Full; available now as
muse-ember-1.0.
Meta Superintelligence Labs has released Muse Ember, the smallest model in its Muse family, for tool calls, routing, extraction, coding subtasks, and long-document parsing. The hosted model is available as muse-ember-1.0 through Muse Code and the Meta Model API.
Meta’s launch post positions Ember as a subagent that can handle hundreds of intermediate steps without sending every request to a larger frontier model. That approach could lower the cost of agent systems, provided Ember maintains sufficient accuracy and does not trigger expensive retries.
The price changes the routing math
| Token type | Price |
|---|---|
| Input | $0.10 |
| Cached input | $0.004 |
| Output | $0.50 |
A request with 100,000 regular input tokens and 10,000 output tokens would cost about $0.015. If all input tokens qualified for the cached rate, the same request would cost about $0.0054. Actual savings depend on cache eligibility, hit rate, output length, and the number of retries.
One million tokens and four input modes
Ember accepts long and multimodal inputs while exposing the interfaces needed for agent workflows:
- A 1-million-token context window
- Up to 512,000 output tokens
- Native text, image, audio, and video input
- Responses API support
- Hosted access through Muse Code and the Meta Model API
Meta claims a lower-cost efficiency frontier
Meta says Ember reaches 36 on the Artificial Analysis Intelligence Index v5 at maximum reasoning, with an estimated cost of $0.14 per task. Claude Haiku 5.5 scores 36.2 at $0.411 per task, giving Ember a similar composite score at roughly one-third of the estimated cost. Meta also says Ember exceeds every GPT-6 Luna reasoning setting on that index.
Maximum reasoning allocates more inference to each answer and generally raises quality, latency, and cost. Artificial Analysis task costs are benchmark estimates based on its prompts and measured token use; production bills will vary with workload shape.
Agent benchmarks favor Ember
| Benchmark | What it measures | Muse Ember | Claude Haiku 5.5 | GPT-6 Luna |
|---|---|---|---|---|
| AutomationBench-AA | Agent automation | 37.8% | 13.9% | 20.2% |
| Terminal-Bench 4.0 | Terminal-based tasks | 40.9% | 39.2% | 16.4% |
| OSWorld 2.1 | Computer use across desktop applications | 69.5% | 68.9% | 50.5% |
| FrontierCode 2.0-AA | Mergeable software-engineering work | 27.2% | 34.9% | 25.2% |
AutomationBench-AA results use each model’s maximum reasoning setting. Terminal-Bench 4.0 uses each model’s best published harness, so the score reflects both the model and its surrounding agent setup. OSWorld 2.1 includes partial credit for progress through desktop tasks.
OSWorld is especially relevant to agents that must inspect screenshots, read interfaces, and operate applications. Ember’s 69.5% score places it close to Haiku 5.5 and well above Luna in Meta’s comparison.
Merge-ready coding remains the weak spot
FrontierCode 2.0-AA gives Claude Haiku 5.5 a clear lead at 34.9%, compared with Ember’s 27.2%. Teams assigning complex, multi-step pull requests should retain a stronger coding model for planning, review, or final implementation.
Meta recommends pairing Ember with the larger Muse Wonder model. Wonder handles planning and difficult decisions; Ember executes the repeated, high-volume steps.
A frontier model trained the specialist
Meta says Ember is its first model trained with Muse Wonder Full as the teacher. This form of distillation uses a larger model to generate or evaluate training signals for a smaller specialist, concentrating useful behavior in a cheaper model that can serve more requests.
Under Meta’s Superintelligence Scaling Framework, Ember remained below the company’s high-risk thresholds for chemical and biological threats, cybersecurity, and loss-of-control scenarios. Meta also applied its core safety training and tested resistance to prompt injections embedded in web pages, documents, and tool output.
Partners report lower latency and shorter answers
Cloudflare’s chief technology officer reports that Ember delivered tail latency up to 30% lower than the comparable model the company tested and performed well on broad investigations spanning multiple repositories.
You.com’s inference lead measured 2.2 seconds at the 95th percentile on coding tasks, with 36% to 48% fewer output tokens and latency four to 13 times lower than leading open-weight models. Shorter responses directly reduce spending because Ember’s output tokens cost five times as much as regular input tokens.
These partner-reported figures use company-specific workloads, harnesses, and infrastructure. Production evaluations should reproduce the measurements with representative prompts, tool schemas, concurrency, and output limits.
Three models, three routing roles
| Model | Suggested role | Estimated AA task cost |
|---|---|---|
| Muse Wonder | Planning and difficult problems | Not stated |
| Muse Spark 1.3 | Coding and agent workflows | $2.39 |
| Muse Ember | Repeatable subagent tasks at scale | $0.14 |
Run the swap test on real traces
Teams currently routing subagent work to Haiku or Luna can evaluate Ember with a controlled replay of production traces:
- Measure end-to-end task completion, tool-call accuracy, and retry rates.
- Track median and 95th-percentile latency under realistic concurrency.
- Record input, cached input, and output tokens separately.
- Test prompt-injection defenses with untrusted documents and tool results.
- Keep a stronger model available for merge-ready code and failed subagent runs.
Ember’s strongest case is repetitive agent work with reusable context, frequent tool calls, or multimodal computer use. Its low cached-input price can produce substantial savings, but total cost depends on cache hits, answer length, and whether lower benchmark performance creates additional retries.
Comments
