Muse Ember: Meta's $0.10/$0.50 small model ties Haiku 5.5 at 36, no weights announced

Meta positions Ember for the work agents repeat all day: tool calls, coding subtasks, routing, extraction and long-document parsing. It is the first model Meta trained using Muse Wonder Full. Meta's portfolio assigns Muse Wonder to the hardest problems, Muse Spark to coding and agentic workflows, and Ember to repeatable coding and agentic tasks at scale. For the hardest problems, Meta suggests pairing them: Wonder plans and Ember handles the volume.

The model ID is muse-ember-1.0. It is available on Muse Code and the Meta Model API, supports the Responses API, and is also available on Google Cloud and Oracle Cloud. It natively understands text, images, audio and video. It has a 1M-token context window and up to 512K output tokens.

The announcement does not give a parameter count, an architecture, weight availability, a licence, quantization options or any memory or hardware requirement. Engineers looking for a small model to run on-device or on their own GPUs have nothing to size it against. One commenter on Hacker News asked Meta to release the models for local use, and the release does not address local availability. Meta says only that it has even more efficient models in training.

Against Meta's own Muse Spark 1.3, Ember gives up one index point (36 vs 37) for roughly one-seventeenth of the cost per task ($0.14 vs $2.39). Its list prices are 12.5 times lower on input and 8.5 times lower on output.

The 200K threshold matters for anyone planning to use the 1M window to load a full codebase or document set in one call, which Meta describes as a use case. Past 200K tokens, input, cached input and output all cost three times as much. Meta's per-task figures also need care. The chart caption calls Ember's costs estimates at proposed pricing, while the note under the chart and the footnote say list pricing.

Meta's chart plots each reasoning-effort setting against Artificial Analysis cost per task, from Intelligence Index v5 as published November 9, 2026. Ember's two plotted points are 33.9 at $0.116 and 35.8 at $0.136. Claude Haiku 5.5 runs from 24.6 at $0.092 to 36.2 at $0.411, passing 34.3 at $0.248, while GPT-6 Luna peaks at 27.8 for $0.152. Those points support Meta's claim that Ember matches Haiku's best at about a third of the cost and beats every Luna setting. At max effort, Meta's table reports these results where Ember leads:

Two caveats apply. Agents' Last Exam, the Personal Superintelligence Index, OSWorld 2.1 and PerceptionBench are Meta's own evaluations, not Artificial Analysis runs. Terminal-Bench 4.0 runs each model in its best published harness: mini-swe-agent for Ember, Claude Code for Haiku 5.5 and Codex for Luna. That makes it a harness-plus-model comparison. One commenter on Hacker News also noted that Luna 6 is about two months old and was only a slight step above Luna 5.6, so a newer small model beating it was no surprise to them.

Ember does not win everywhere. On FrontierCode 2.0-AA, which judges real engineering tasks on mergeability, Haiku 5.5 scores 34.9 against Ember's 27.2 and Luna's 25.2. On GDP.pdf-AA, which covers professional work over long documents, Ember and Luna tie at 22.8, with Haiku at 20.8. On OSWorld 2.1, Ember leads on both partial-credit reward (69.5 vs Haiku's 68.9 and Luna's 50.5) and binary scoring (33.8 vs 27.8 and 17.0), though Meta's chart uses the partial-credit numbers. For computer-use agents where only full task completion counts, the binary scores are the ones to plan around, and they are less than half the partial-credit figures.

For code that has to merge without rework, Haiku 5.5 is ahead. Meta frames Ember's coding role as the subtasks an agent sends off hundreds of times a session: reading files, running tests, making targeted fixes. The Cline founder described it as suited to

One commenter on Hacker News believes Ember is the stealth model previously served on OpenRouter as Crouton Alpha. Another, who used Crouton Alpha heavily, said it starts strong but is quite bad by turn 10 and unusable by turn 20. Meta has confirmed neither claim. Meta's evidence covers single long inputs, such as the 1M window and GDP.pdf-AA, plus end-to-end task scores, and it reports nothing on degradation across many conversational turns. Teams using Ember as a subagent with short, scoped calls are closer to its stated design than teams running it as a long-lived primary agent.

Reasoning effort is a separate open question. A commenter counted six effort levels (minimal, low, medium, high, xhigh, max) and found that higher effort did not reliably produce better output on a single SVG drawing test. Another said only the max setting, after a 27-second wait, produced a drawing with the parts placed correctly, while a third called one sample too thin to measure capability. One commenter called so many reasoning levels an antipattern for users who have to choose between them. Meta's chart plots only two Ember settings and publishes no guidance on which effort level suits which task type.

Meta's release has no first-party latency or throughput figures, so the speed claims come from partner testing. You.com reports a p95 of 2.2 seconds on coding quests, about 4x to 13x faster than leading open-weight models, with 36 to 48 percent fewer output tokens. Cloudflare reports up to 30 percent faster tail latency than the comparable model it tested, which it did not name. Neither partner gives hardware, region, prompt sizes or the full set of comparison models. Box, Cline and Oracle offered qualitative statements without numbers, though Box says it tested Ember on its Box AI Complex Work Eval.

On safety, Meta says it evaluated Ember under its Meta Superintelligence Scaling Framework, covering chemical and biological, cybersecurity and loss-of-control risks, and found it remained below the framework's high-risk thresholds. Meta also evaluated whether Ember resists instructions hidden in web pages, documents or tool outputs and whether it seeks human input before high-impact actions. Meta publishes no prompt-injection success rates or other scores from those tests, so agent builders should keep their own guards on tool output.