Meta's Muse Ember matches Haiku 5.5 and GPT-6 Luna on list price and competes on tokens per task

On November 9, Meta Superintelligence Labs released Muse Ember, which it calls its smallest and most efficient model. It is available as muse-ember-1.0 on the Meta Model API and in Muse Code, at $0.10 per million input tokens and $0.50 per million output tokens, with a 1M-token context window and up to 512K output tokens. Meta aims it at the steps agents repeat all day: tool calls, coding subtasks, routing, extraction and long document parsing.

The headline price is not new. Claude Haiku 5.5 and GPT-6 Luna already charge $0.10 and $0.50. What Meta is claiming is that Ember finishes the same work with far fewer tokens, so the bill per task is lower even though the rate card is identical. That makes this launch a useful test of how much list prices still tell you.

Same rate card, a third of the bill

Meta’s central chart plots the Artificial Analysis Intelligence Index (v5, as published November 9) against cost per task at each reasoning-effort setting:

Model	Setting	Index score	Cost per task
Muse Ember	xhigh	33.9	$0.116
Muse Ember	max	35.8	$0.136
Claude Haiku 5.5	xhigh	34.3	$0.248
Claude Haiku 5.5	max	36.2	$0.411
GPT-6 Luna	max	27.8	$0.152

At maximum effort, Ember scores 35.8 for $0.136 per task and Haiku scores 36.2 for $0.411. Meta’s summary table rounds both to 36 and says Ember “matches Haiku’s best score at about a third of the cost.” Strictly, Haiku is 0.4 points ahead and three times the price. Against Haiku at xhigh, Ember is 1.5 points better at 55% of the cost. Luna tops out eight points lower at a slightly higher cost than Ember.

Because all three models share the same standard token prices, the gap in cost per task has to come almost entirely from how many tokens each one spends, mostly reasoning and output tokens. A threefold cost difference at equal rates implies roughly a threefold difference in token use on this benchmark. One of Meta’s launch partners gives a figure in the same direction: You.com’s inference lead says Ember used 36-48% fewer output tokens than the leading open-weight models it tested, with a p95 latency of 2.2 seconds.

Two details qualify the comparison. The chart caption says Ember’s costs are “estimates at proposed pricing,” while the footnote says list pricing, so the Ember points appear to be computed by Meta from Artificial Analysis token counts and should be rechecked when the index lists the model independently. And the standard rate only applies to prompts up to 200K tokens. Above that, Ember costs $0.30 input and $1.50 output, three times as much, which matters for a model advertised as able to take a whole codebase in one call. Cached input is $0.004 per million, a 96% discount, which rewards agents that keep a stable prefix.

Where Ember leads, and where it does not

Meta publishes ten evaluations at max effort. The spread across them is more informative than the index:

Benchmark	Muse Ember	Haiku 5.5	GPT-6 Luna	Source
AutomationBench-AA	37.8	13.9	20.2	Artificial Analysis
Coding Agent Index-AA	46	37	31	Artificial Analysis
FrontierCode 2.0-AA	27.2	34.9	25.2	Artificial Analysis
GDP.pdf-AA	22.8	20.8	22.8	Artificial Analysis
Terminal-Bench 4.0	40.9	39.2	16.4	Best published harness
Agents’ Last Exam	61.6	57.0	57.1	Meta internal
OSWorld 2.1 (binary)	33.8	27.8	17.0	Meta internal
PerceptionBench (with tools)	80.6	78.6	62.8	Meta internal

The largest lead is AutomationBench, which scores end-to-end business workflows on strict pass@1: Ember completes 37.8% against 13.9% for Haiku. That is the closest match to the model’s stated purpose of pulling records, calling tools in sequence and updating systems.

The clear loss is FrontierCode 2.0, which judges real engineering tasks on whether the result is mergeable. Haiku scores 34.9% to Ember’s 27.2%. Read alongside the Coding Agent Index win, the pattern suggests Ember is strong at the mechanics of working inside a harness and weaker at producing a change a maintainer would accept. Meta’s own positioning agrees: it recommends pairing Ember with Muse Wonder, with Wonder planning and Ember handling volume.

Several results need care. Terminal-Bench uses each model’s “best published harness,” which means mini-swe-agent for Ember, Claude Code for Haiku and Codex for Luna, so the 1.7-point lead over Haiku mixes model and scaffold. On OSWorld, Meta leads with a partial-credit score of 69.5% against Haiku’s 68.9%. The binary pass rates are 33.8% and 27.8%, and those are the ones that say whether the task was done. Agents’ Last Exam was run on a Linux subset. The Personal Superintelligence Index, where Ember leads by ten points, is Meta’s own composite of eight evaluations.

Trained with Muse Wonder Full, with no word on weights

Meta says Ember is “the first trained using Muse Wonder Full,” the larger model it introduced with Muse Wonder in October and described then as a tool for building specialized models. The post does not say what that training involved, though using a large model to generate training signal for a small one is the standard way to get this kind of efficiency. It gives no parameter count and no architecture details.

Ember is natively multimodal across text, images, audio and video, and supports the Responses API format, which lowers the switching cost for teams already built on OpenAI-style tooling. It is also listed on Google Cloud and Oracle Cloud alongside Spark and Wonder. The post makes no mention of open weights. Meta has published weights for other Muse models, including Muse Glimmer and a version of Muse Spark, so the omission is worth noting for teams that wanted a small model to run themselves.

On safety, Meta says Ember was evaluated under its Superintelligence Scaling Framework for chemical and biological, cybersecurity and loss-of-control risks and stayed below the high-risk thresholds, and that it was tested on agent behaviors such as resisting instructions hidden in web pages and tool outputs. It links to a separate evaluation report and publishes no numbers for those tests in the post.

What it means for agent builders

The practical use is as a subagent. Meta’s comparison with its own Muse Spark 1.3 makes the case: Ember scores 36 on the index to Spark’s 37, at $0.14 per task against $2.39, about one seventeenth of the cost. If those figures hold on real workloads, routing file reads, test runs, extraction and narrow fixes to Ember and keeping a larger model for planning and review is an easy saving. Cline’s founder describes that pattern in Meta’s partner quotes, and Cloudflare’s CTO reports tail latency up to 30% better than a comparable model.

The broader point is about how to compare models now that three vendors charge the same rates. Price per million tokens has stopped being a useful number at this tier. Tokens per completed task, and the pass rate on your own tasks, decide the bill. That also means vendor charts built on a benchmark’s token counts will not transfer cleanly: a model that is frugal on index questions may not be frugal in your harness with your tool outputs.

Before moving traffic, run your own task set at two effort settings and record total tokens and pass rate for each model. Check how much of your traffic exceeds 200K-token prompts. And wait for Artificial Analysis to publish Ember’s results directly, along with any independent FrontierCode or Terminal-Bench runs on a common harness. Meta says more efficient models are already in training, so Anthropic and OpenAI now have a reason to answer on efficiency settings as well as price.