Meta launches Muse Ember, a low-cost model for high-volume agent work
Meta says its smallest Muse model matches Claude Haiku 5.5 on a key index at lower cost. Artificial Analysis, which runs the index, reports the same score but finds knowledge gaps.
Key takeaways
- Meta released Muse Ember on November 9 through its Model API and Muse Code, describing it as its smallest and most efficient model.
- Meta charges $0.10 per million tokens of input and $0.50 per million of output, with higher prices for prompts above 200k tokens.
- Artificial Analysis reports an Intelligence Index score of 36 at about $0.14 per task, matching Claude Haiku 5.5 at max reasoning.
- Artificial Analysis found mixed coding results and weak knowledge accuracy, with the model declining 64% of questions in one test.
- Some benchmark figures in Meta's charts differ from those published by Artificial Analysis, and no source explains the gap.
Meta has released Muse Ember, which it calls its "smallest and most efficient model," aimed at the repetitive steps AI agents carry out, such as tool calls, coding subtasks, routing, data extraction and parsing long documents. The model went on sale on November 9 through Meta's Model API and its Muse Code coding agent under the identifier muse-ember-1.0, according to Meta's announcement. Meta's announcement does not give a parameter count, and neither the announcement nor the independent coverage reviewed by NextFish describes any release of model weights or an open licence.
What does Meta say Muse Ember can do?
Meta lists a context window of 1M tokens and up to 512K output tokens, and says the model natively handles text, images, audio and video. The company says it can work through long documents and operate a computer from screenshots. Meta also says Muse Ember is the first model it trained using Muse Wonder Full, its larger system, though SiliconANGLE notes Meta did not say which part of the training process the larger model was used for.
Meta positions the model as a worker paired with its bigger models. In the company's words, "Muse Wonder plans, Muse Ember handles the volume."
On safety, Meta says it tested the model under its Meta Superintelligence Scaling Framework, covering chemical and biological, cybersecurity and loss-of-control risks, and that it stayed below the framework's high-risk thresholds. These are Meta's own findings.
How much does it cost?
Meta prices input at $0.10 per million tokens, cached input at $0.004 per million and output at $0.50 per million. A developer post from Meta adds that prompts above 200k input tokens cost more, at $0.30 input and $1.50 output per million tokens.
Developer Simon Willison wrote that for shorter prompts the input and output prices are the same as OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5, though Ember's cached input is cheaper ($0.004 against $0.01 per million), and that all three raise prices past different prompt-length thresholds. Taking those into account, he places Ember between the other two. He also argued that per-token prices are a weak comparison because reasoning models can use very different numbers of tokens on the same problem.
How does it compare on benchmarks?
Meta's headline claim is a score of 36 on the Artificial Analysis Intelligence Index for Muse Ember, level with Claude Haiku 5.5 at its highest reasoning setting after rounding (Meta's chart gives 35.8 against 36.2), for an estimated $0.14 per task (Meta's developer post says about $0.13). Artificial Analysis, which runs the index, published its own write-up reporting the same score and cost (Meta's chart is drawn from Artificial Analysis's results), and said Claude Haiku 5.5 (max) costs $0.41 per task for that score. It placed Muse Ember on its Pareto frontier of score against cost per task, and attributed the lower cost mainly to the cheaper cached input, which it says makes up most of the cost in its agentic tests, and to Ember using roughly a third as many output tokens as Haiku.
The independent results were not uniformly favorable. Artificial Analysis said Muse Ember beat Claude Haiku 5.5 on Terminal-Bench 4.0 but trailed it on FrontierCode 2.0 (27.2% vs 34.9%), calling agentic coding "mixed." It named knowledge as the main weakness: an AA-Omniscience accuracy of 21.1%, against 34.7% for Claude Haiku 5.5 (max), the lowest of the similarly capable models it compared. The firm said this is largely because the model declines to answer 64% of questions, which also keeps its hallucination rate at 17.4%, half of Haiku's by the firm's measure.
Some figures differ between Meta's charts and the Artificial Analysis article. Meta's chart shows Muse Ember at 37.8% on AutomationBench-AA, while Artificial Analysis reports 68.1%. For Claude Haiku 5.5 on Terminal-Bench 4.0, Meta shows 39.2%, noting each model ran in its "best published harness," while Artificial Analysis reports 32.8%. Artificial Analysis says its Terminal-Bench figure comes from its own Intelligence Index evaluations, and Meta's AutomationBench chart is itself credited to Artificial Analysis and labeled "strict pass@1." Neither source explains the gap, and NextFish has not independently verified the results. Several of Meta's other benchmarks, including OSWorld 2.1 and PerceptionBench, are labeled by Meta as internal evaluations.
In informal testing, Willison found that the model's six reasoning levels did not map cleanly to output quality in his drawing test, and that one level failed to return a complete result.
Who has tried it?
Meta quoted early partners including Box, Cline, Oracle, Cloudflare and You.com. Cloudflare CTO Dane Knecht said the model ran "up to 30% faster at the tail than the comparable model we tested." These are partner statements published by Meta.
Disclosure: NextFish's AI drafting desk uses Claude, a model made by Anthropic, whose Claude Haiku 5.5 is compared in this story.
Questions readers are asking
Is Muse Ember an open-weight model?
Meta's announcement describes availability through its Model API and Muse Code. Neither Meta's post nor the independent coverage reviewed mentions a weights release or an open licence.
How much does Muse Ember cost?
For prompts up to 200k tokens, Meta charges $0.10 per million tokens of input, $0.004 per million of cached input and $0.50 per million of output, with higher rates above that.
Has anyone tested Meta's benchmark claims?
Artificial Analysis, whose index Meta cites, published its own write-up reporting a score of 36 but reporting mixed coding results and low knowledge accuracy. Some of its figures differ from Meta's charts.
What is Muse Ember designed for?
Meta says it targets repetitive agent work such as tool calls, routing, coding subtasks, data extraction and parsing long documents, often paired with its larger Muse Wonder model.
Sources
- Meta: Introducing Our Smallest, Most Efficient Model: Muse Ember
- Meta: Start building with Muse Ember, our smallest, most efficient model
- Artificial Analysis: Meta's Muse Ember is the cheapest model scoring 36 on the Intelligence Index
- SiliconANGLE: Meta introduces Muse Ember to power cost-efficient AI agents
- Simon Willison's Weblog: Introducing Our Smallest, Most Efficient Model: Muse Ember (link post)
- #meta
- #muse-ember
- #small-models
- #ai-agents
- #benchmarks
- #model-pricing
Get the day's top AI & tech stories on WhatsApp: Follow the NextFish channel
Get stories like this every morning.
Related
AI2 min
HUMAIN sets Dec. 15 for its AI stack, promises a 3T-parameter model
The Saudi AI company's CEO calls the unnamed model the world's largest open-weight model and says it is far cheaper to run. No model card, licence or benchmark has been published.
Agents3 min
OpenAI launches ChatGPT Wallet so its dots agents can buy for users
The payment feature starts with US ChatGPT Pro subscribers. Users approve each purchase with a biometric check, and agents get a payment token, not the real card number.
Big Tech3 min
Italy fines Meta €805 million over Facebook's phone gallery access
The Garante says the Facebook app for Android accessed photos and videos on users' phones without specific consent. Meta has 60 days to change how the feature asks for permission.
AnalysisAgents3 min
How big shopping and travel sites now write rules for AI agents
eBay, Target, Booking, Delta and Shopify have all added terms about AI agents. NextFish compared their terms with archived copies to see what changed and when.