Muse Ember: Meta's $0.10 Subagent Model vs Haiku 5.5 and Luna

Last updated: November 10, 2026

Muse Ember is Meta's new small model, and its price is the story: $0.10 per million input tokens and $0.50 per million output, which is exactly the list price of Claude Haiku 5.5 and GPT-6 Luna. Meta pitches it as the model for "the work agents repeat all day": tool calls, coding subtasks, routing, extraction and long-document parsing. If you run a harness with a cheap worker tier, this is a third candidate for that slot, and the sticker price will not tell you which to pick.

Official Sources#

What shipped#

Per Meta's announcement, Muse Ember is available today on the Meta Model API and Muse Code as muse-ember-1.0. It has a 1M-token context window, up to 512K output tokens, native text, image, audio and video input, and supports the Responses API. Meta says it is the first model trained using Muse Wonder Full, and positions it as the volume tier: Muse Wonder for the hardest problems, Muse Spark for coding and agentic workflows, Ember for repeatable tasks at scale.

There are no open weights. Unlike Muse Glimmer, you cannot run this one locally, and several HN commenters said that was their first reaction.

Pricing next to the two models it targets#

Meta's figures come from its own page; the Haiku and Luna rows are from our Haiku vs Luna subagent analysis, which checked each vendor's pricing page. Two things differ. First, Ember's cached input is 4% of its input price, against 10% for the other two, a point also raised on HN. In an agent loop that resends a growing prompt every turn, cache reads dominate the bill, so this is the line that matters. Second, Haiku's rates jump once a prompt passes 100K tokens (cache reads go from $0.01 to $0.05 in that analysis). Meta's page lists one set of rates and no long-prompt tier. We did not find a pricing page that confirms there is none, so check the Meta API docs before you assume a long loop stays at $0.004.

Benchmarks, labelled#

Everything below is from Meta's page. Artificial Analysis (AA) rows are AA measurements that Meta reproduces; "Meta internal evaluation" rows are vendor-reported and unaudited.

Two readings stand out. On the headline index Ember ties Haiku 5.5 at 36, and Meta's chart puts Ember's cost per task at about $0.14 against $0.41 for Haiku at the same score, which is where the "a third of the cost" claim comes from. But Haiku beats Ember on FrontierCode, the benchmark that judges mergeability of real engineering tasks. One HN commenter noted that this kind of test is harder to game than the others, and read the gap as a warning. Meta itself says to pair Ember with Muse Wonder for the hardest problems.

The cost-per-task spread is also wide within each model because it depends on the reasoning setting. In Meta's chart Haiku 5.5 runs from roughly $0.09 to $0.41 per task across its four settings, and Luna from about $0.07 to $0.15. Ember's two plotted settings sit at $0.116 and $0.136, and Meta labels Ember's costs as estimates at proposed pricing. Compare at the effort level you will actually run, not at the max.

Run it#

Meta's page points developers to dev.meta.ai and says the model supports the Responses API under the id muse-ember-1.0. We have not verified an OpenCode provider entry for it, so do not paste a guessed opencode run line into your config; check your harness's model list first. If you already route Muse models through Claude Code or Codex, our Meta Muse setup guide covers the working paths for the earlier Muse models.

A low-risk way to test it is shadow mode: send your worker traffic to both your current cheap model and Ember, apply only the incumbent's result, and diff the two. A bounded subagent task such as summarising a diff, extracting fields or routing a ticket is the right shape for this. Hard multi-file edits are not.

Which one to pick#

What people are actually saying#

The HN thread is dominated by a pelican-SVG test across Ember's six effort levels, which says little about agent work. The substantive comments split two ways. Supporters like the price and the efficiency numbers: one commenter compiled the AA figures and called it a strong all-round value model, and another who uses Haiku and Luna for support work on software tasks welcomed an even cheaper option. Skeptics say Meta's Muse models read as frontier on benchmark cards but feel weaker in real use, point out there are no weights, and note a US-hosted open-weight DeepSeek variant is cheaper on OpenRouter. One user who had used Muse Spark for documentation work said Haiku 5.5 has since outclassed it at that size. None of this is evidence either way; it is the reason to run your own shadow test.

FAQ#

What is the API model id for Muse Ember?#

muse-ember-1.0, per Meta's launch post. It is available on the Meta Model API and Muse Code and supports the Responses API.

How much does Muse Ember cost?#

$0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens, with a 1M-token context and up to 512K output tokens.

Are Muse Ember weights open?#

No. Meta lists it as API and Muse Code only, and HN commenters confirmed no weights were released.

Sources#

Continue Reading#

• Haiku 5.5 or GPT-6 Luna as a subagent - the price-cliff math for the two models Ember targets
• Claude Haiku 5.5 release guide - tiers, tokenizer change and migration notes
• Meta Muse Spark 1.1 developer guide - the pricier Muse tier Ember sits under
• Muse Glimmer 30B open weights - the Meta model you can actually run locally
• Claude Code and Codex with Meta Muse - setup paths that work for earlier Muse models
• AI API Cost Calculator - estimate per-request and monthly cost before you pick a subagent model
• AI API Pricing Comparison - per-token input and output prices side by side