Meta releases Muse Ember at 10 cents per million input tokens for the steps agents repeat
The Monday post prices the model at $0.10 in and $0.50 out, with a 1 million token context window, and says it is the first model trained with Muse Wonder Full. On the Artificial Analysis index Meta reprints, Ember and Claude Haiku 5.5 both land at 36. On FrontierCode, Haiku is ahead. A caption calls the Ember costs estimates, and another line on the same chart says list pricing.
Feature3 minIn-depth
Priya Shah3 min read
- Meta
- Muse
- Agents
- Pricing

Photo: MetaSource page
Meta on Monday introduced Muse Ember, which it calls its smallest and most efficient model, built for the work it says agents repeat: tool calls, coding subtasks, routing, extraction, and long-document parsing. The post is dated 9 November. Ember is available as muse-ember-1.0 on the Meta Model API and in Muse Code. The list price in the availability section is $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens, for prompts up to 200,000 tokens. Above that, the footnote says, the rates are $0.30, $0.012, and $1.50. The context window is 1 million tokens, and the maximum output is 512,000. The post says Ember, Muse Spark, and Muse Wonder are now on Google Cloud and Oracle Cloud.
The efficiency claim is Meta’s reading of an Artificial Analysis chart dated 9 November. The prose says Ember delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting, that it matches Haiku’s best score at about a third of the cost, and that it beats every Luna setting. The table underneath lists an Artificial Analysis index of 36 for Ember at max and 36 for Haiku at max, against 28 for Luna at max. One caption on the chart says Ember’s costs are estimates at proposed pricing. Another line, under the same chart, says the cost is at list pricing. The availability section is where the dollar rates are stated as the price on offer.
Where the chart has Ember ahead, and where it does not
The comparison is not uniform, and the post’s own table shows that. On FrontierCode 2.0, as run for Artificial Analysis, Haiku at max is listed at 34.9 percent and Ember at 27.2 percent, with Luna at 25.2 percent. On Terminal-Bench 4.0 the order flips: Ember 40.9 percent, Haiku 39.2 percent, Luna 16.4 percent. The footnote says those three numbers use each model’s best published harness, and it names them: mini-swe-agent for Ember, Claude Code for Haiku, and Codex for Luna. They are not the same scaffold. On AutomationBench, which Meta credits to Artificial Analysis as strict pass@1, Ember at max is 37.8 percent, Luna 20.2 percent, and Haiku 13.9 percent. On a Meta-internal composite it calls the Personal Superintelligence Index, the scores are 46.8, 37.0, and 36.1. The post says that index mixes eight evaluations. It is not the Artificial Analysis number.
A second comparison, with Muse Spark 1.3, is the one Meta uses to place Ember in its own lineup. Spark’s index is 37 against Ember’s 36. The cost per index task is $0.14 for Ember and $2.39 for Spark. Spark’s token price in that table is $1.25 per million input and $4.25 per million output. The sentence Meta attaches is that Wonder is for the hardest problems, Spark for coding and agentic workflows, and Ember for repeatable coding and agentic tasks. It also says Ember is the first model trained using Muse Wonder Full.
What stood out in our testing was speed. Muse Ember ran up to 30% faster at the tail than the comparable model we tested, and it handled broad, multi-repo investigation tasks well.
— Dane Knecht, chief technology officer, Cloudflare, in Meta’s post
Knecht does not name the model Ember was faster than. Yashodha Bhavnani, vice president of AI products at Box, is quoted on a contract scored against a negotiation playbook and on candidate screening. Saoud Rizwan, founder of Cline, is quoted on using it as a subagent. Those are early-partner comments in the launch post, not a public leaderboard.
On safety, Meta says Ember was evaluated under its Superintelligence Scaling Framework, including chemical and biological risk, cybersecurity, and loss of control, and that the evaluations found it remained below the framework’s high-risk thresholds. The post also says it was tested on whether it follows user intent, handles sensitive information, resists instructions hidden in pages, documents, or tool outputs, and asks a person before high-impact actions. It does not publish the scores for those tests. The price, the index tie with Haiku at 36, and the FrontierCode gap the other way are the figures a buyer can check against the tables on the page.


