Summary: Meta released Muse Ember (muse-ember-1.0) on November 9, 2026, a small multimodal model built for the high-volume steps agents repeat: tool calls, coding subtasks, routing, extraction, and long document parsing. It launches at $0.10 per million input tokens and $0.50 per million output, and Meta claims it sits on the efficiency frontier of the Artificial Analysis leaderboard. For anyone running agent workflows, this changes which model you route the cheap, frequent calls to.
What actually shipped
Muse Ember is available today on Muse Code and the Meta Model API under the name muse-ember-1.0, with support for the Responses API. It takes a 1M-token context window and produces up to 512K output tokens, and it natively handles text, images, audio, and video, so one model can read screenshots, parse scanned documents, or work from meeting recordings without a separate vision pipeline.
The positioning is explicit. According to Meta’s announcement, the model targets “complex reasoning execution” — the structured, mid-difficulty steps beneath a planner. Meta frames it as the bottom tier of a three-model portfolio: Muse Wonder for the hardest problems, Muse Spark for coding and agentic workflows, and Muse Ember for repeatable tasks at scale. Where the pairing matters, Meta’s guidance is that Muse Wonder plans while Muse Ember handles the volume. I covered how Wonder’s specialization strategy works in an earlier post on this blog, and Ember is notable for another reason: it is the first model trained using Muse Wonder Full.
The model offers six reasoning levels — minimal, low, medium, high, xhigh, and max — which matters for cost control, since you can dial effort per call rather than per deployment.
The pricing, and where it jumps
At list price, Ember costs $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens. That matches the short-context pricing of Claude Haiku 5.5 and GPT-6 Luna, as Simon Willison points out in his write-up of the launch.
The catch is the long-context threshold. All three models raise prices past a prompt-size cutoff, and Ember’s comes the earliest:
| Model | Prompt size | Input $/M | Output $/M |
|---|---|---|---|
| GPT-6 Luna | ≤272k / >272k | $0.10 / $0.20 | $0.50 / $0.75 |
| Muse Ember 1.0 | ≤200k / >200k | $0.10 / $0.30 | $0.50 / $1.50 |
| Claude Haiku 5.5 | ≤100k / >100k | $0.10 / $0.50 | $0.50 / $2.50 |
Past 200k tokens, Ember costs more per input token than Luna, and three times its short-context output rate. If your agent stuffs full codebases or large document sets into single calls, run the math at your actual prompt sizes before switching.
What the numbers say — and where they come from
Meta’s headline claim is cost-per-task, not cost-per-token, which is the right frame now that reasoning tokens vary so much between models. On the Artificial Analysis Intelligence Index, Meta’s chart shows Ember scoring 33.9 at $0.116 per task and 35.8 at $0.136, versus Claude Haiku 5.5 reaching 36.2 at $0.411 and GPT-6 Luna topping out at 27.8 at $0.152. One caveat stated plainly in the source: Ember’s costs in that chart are “estimates at proposed pricing,” and several evaluations — Agents’ Last Exam, the Personal Superintelligence Index, OSWorld 2.1, PerceptionBench — are Meta internal evaluations, not independent audits.
The benchmark picture is mixed in a useful way. Ember wins AutomationBench-AA (37.8 vs Luna’s 20.2 and Haiku’s 13.9), Agents’ Last Exam (61.6), Terminal-Bench 4.0 (40.9), Coding Agent Index-AA (46), and PerceptionBench (80.6). It loses FrontierCode 2.0-AA to Haiku, 27.2 against 34.9, on real engineering tasks judged on mergeability. Against Meta’s own Muse Spark 1.3, Ember’s Artificial Analysis Index is 36 versus 37 — but at roughly $0.14 per task versus $2.39. The trade is clear: a few points of peak capability for an order of magnitude less spend.
Partner quotes fill in latency. Dane Knecht, Cloudflare’s CTO, reported Ember ran up to 30% faster at the tail than the comparable model they tested. Chak Pothina, You.com’s inference lead, reported a p95 of 2.2 seconds — 4x to 13x faster than leading open-weight models, with 36–48% fewer output tokens. These are early partner tests, not benchmarks you can reproduce from the blog post.
Where the material runs out
The excerpt of Meta’s blog I worked from cuts off mid-sentence in the closing section, so the final paragraphs of the announcement are missing here — nothing above should be read as summarizing the page’s ending. The sources also do not establish independent verification of the internal evaluations, real-world long-context costs at your prompt distribution, or open-weights availability; the launch post describes API access only.
What to check next
If you run agent harnesses, the concrete move is a routing test: point your highest-volume subagent calls — tool invocation, extraction, file reading — at muse-ember-1.0 at medium or high reasoning, and compare cost per completed task against whatever you use today. Watch two things. First, whether quality holds on mergeability-style tasks, where FrontierCode suggests Ember trails Haiku. Second, your prompt-size distribution: if calls routinely exceed 200k tokens, the threshold pricing can erase the savings. Simon Willison’s hands-on note adds a soft signal worth noting — his test outputs at max reasoning were comparable to what he got from Muse Wonder, a model he prices at 40x Ember’s cost — but that is one informal test, so treat your own workload as the real benchmark.