muse ember: Meta's $0.10 Agent Model Claims Haiku-Level Scores at a Third of the Cost
The muse ember model is Meta's new small agent model, out Nov 9 at $0.10/$0.50 per million tokens with a 1M context; Meta says it matches Claude Haiku 5.5's best score at about a third of the cost per task.
- Context window of 1M tokens and up to 512K output tokens; cached input costs $0.004 per million tokens.
- Available now on the Meta Model API and in Muse Code as
muse-ember-1.0, with Responses API support. - TechCrunch says it is the fourth model under the Muse brand, after Spark in April, Glimmer in August and Wonder in October.
in this block
Meta's new small model for AI agents, muse ember, arrived on November 9 from Meta Superintelligence Labs at $0.10 per million input tokens and $0.50 per million output tokens. Meta says it matches the best score of Anthropic's Claude Haiku 5.5 at about a third of the cost per task and beats every reasoning setting of OpenAI's GPT-6 Luna, using results from Artificial Analysis.
What actually happened
Meta's launch post describes the model as built for "the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing." The pitch is cost per task, not raw peak intelligence. TechCrunch puts it in one line from the announcement: "Muse Wonder plans, Muse Ember handles the volume." In practice that means a bigger model writes the rules for a job and the small one applies them across thousands of cases.
TechCrunch also notes that Meta trained it to handle text, images and audio, so an agent can read screenshots, charts and scanned documents, or work from voice memos and recordings, without calling a separate model. On price, TechCrunch says it is far cheaper than Haiku and Luna and competitive with open models like DeepSeek V4.1 Flash and GLM 5.3 Flash.
We covered the bigger sibling in muse wonder and the main rival in claude haiku 5.5.
The benchmark claims
These figures come from Meta's own evaluation table, which uses Artificial Analysis runs for several tests. Treat them as vendor claims until others repeat them:
- Artificial Analysis Index: 36 for Ember at max effort, level with Haiku 5.5's best and ahead of GPT-6 Luna at 28.
- Terminal-Bench 4.0 agentic coding: 40.9, versus 39.2 for Haiku and 16.4 for Luna.
- AutomationBench-AA business workflows: 37.8, versus 13.9 for Haiku and 20.2 for Luna.
- OSWorld 2.1 computer use (partial): 69.5, versus 68.9 for Haiku and 50.5 for Luna.
It does not win everywhere. On FrontierCode 2.0-AA, which judges whether real engineering changes would be merged, Meta's own table lists Haiku 5.5 at 34.9 and Ember at 27.2. So for harder code changes, the cheaper model may not be the better one.
The cost chart is the core of the argument. Meta's data shows the new model scoring 35.8 on the index at about $0.136 per task, while Haiku 5.5 reaches 36.2 at about $0.411 per task. That gap is where the "third of the cost" claim comes from.
What we could not check
Meta's cost figures are estimates based on its proposed list pricing, as a note under its own chart says, and the Personal Superintelligence Index in the table is an internal composite that outsiders cannot reproduce. Meta's post lists access through its API and Muse Code, and we saw no mention of open weights, unlike the open Glimmer release. Real-world speed will depend on load once many developers start calling it at the same time.
What partners said
Meta quoted several early testers. Cloudflare's CTO said the model ran up to 30% faster at the tail than a comparable model it tested. You.com said it finished every coding quest in its test with a p95 latency of 2.2 seconds. Box, Cline and Oracle also gave supportive quotes. These are partner testimonials in a launch post, so they are positive by design.
Why it matters
Agent workloads burn tokens on many small steps, and price per step decides whether an agent is affordable at scale. Muse Ember is Meta's bet that a cheap, fast model handling the bulk of those steps is worth more than a slightly smarter one that costs three times as much. It also adds pressure on Anthropic and OpenAI, whose small models are the obvious comparison in Meta's own charts.
What to do as a reader
If you run agents, read the Meta launch post and test Muse Ember on your own tasks, especially the ones where you now use Haiku or Luna. Compare total cost per finished task, not just token prices, and check accuracy on harder coding work where Meta's own numbers show it trailing.
Independent benchmarks will follow in the coming weeks. Until then, the safest view of Muse Ember is "promising and cheap, with self-reported results."
Not financial advice. DYOR, ser.
