Muse Ember: Meta's Volume Model at the Same $0.10 and $0.50 as Haiku 5.5

Meta released Muse Ember on November 9, 2026. It is Meta's smallest Muse model, the first trained using Muse Wonder Full, and it ships at the same $0.10 input / $0.50 output list price as Claude Haiku 5.5 and GPT-6 Luna. The product story is volume work: tool calls, coding subtasks, routing, extraction, and long document parsing, with Muse Wonder or Muse Spark planning and Ember executing.

On Meta's self-reported, allowed launch rows at max reasoning effort, Ember is slightly ahead of Haiku 5.5 and clearly ahead of Luna. The real differentiators at that shared sticker are a cheaper cache ($0.004), a 200K cheap-tier window (twice Haiku's, short of Luna's), 524,288 max output (four times either peer), and a long-context rate that beats Haiku's and loses to Luna's. Default effort is max, so a default call bills more reasoning tokens than a default Haiku or Luna call. It is not a near-frontier model. See Meta's launch post and model docs.

What Muse Ember is

Meta positions the Muse portfolio as Wonder for the hardest problems, Spark for coding and agentic workflows, and Ember for repeatable coding and agentic tasks at scale. Docs repeat the pairing line: the larger model plans, and Muse Ember handles the volume.

Recommended Ember work is bounded and repetitive: sub-agent coding and terminal commands, tool calls inside a larger workflow, routing and classification, and the same task over thousands of images, files, or records. For hard single-shot coding, knowledge-heavy questions, or visual coding, Meta says to use Spark or Wonder instead. Meta's choose-a-model docs also map Claude Code's haiku alias and sub-agent model to muse-ember-1.0 (opus to Wonder, sonnet to Spark 1.3).

Ember is available today on the Meta Model API and Muse Code as muse-ember-1.0, pay-as-you-go and every Muse Code plan, in Standard mode only. Meta also says Muse Ember, Muse Spark, and Muse Wonder are available on Google Cloud and Oracle Cloud; Meta publishes no first-party per-token prices for those clouds, so treat that as availability only.

Benchmarks

All scores below are self-reported by Meta at max reasoning effort, from the November 9, 2026 launch table. Methodology notes live in Meta's evaluation report. Several Haiku gaps are small. Terminal-Bench uses each model's best published harness, so it is not a common-harness comparison. This is a close Haiku-class result at the same list price, not a frontier claim.

Allowed launch rows

Compact table of the same self-reported max-effort scores. Headline reading: ahead on every allowed row, close to Haiku on several, clearer lead on Personal Superintelligence Index and against Luna.

• Agents' Last Exam: Meta internal, ALE v1.0 Linux subset. Five of 105 tasks excluded for every model. Common internal agent loop, not the official product-vs-product leaderboard. Mean task score with partial credit.
• Terminal-Bench 4.0:Ember in mini-swe-agent, Haiku 5.5 as Anthropic's self-reported Claude Code result, Luna as Codex on the Terminal-Bench leaderboard. Haiku gap is 1.7 points.
• OSWorld 2.1: Meta internal runner, 108 tasks, mean per-task partial credit plus a strict full-credit share. Haiku partial gap is 0.6 points.
• PerceptionBench: Meta internal, 3,000 questions, scored with tools, official judge. Haiku gap is 2.0 points.
• Personal Superintelligence Index:Meta's internal composite of eight personal and knowledge-work evaluations. Report the composite only.

Pricing

The request's input-token count, cached tokens included, sets the Ember tier. A continued conversation via previous_response_id counts as input. Up to 200K input tokens: cached input $0.004, input $0.10, output $0.50. Over 200K, the whole requestis billed at cached $0.012, input $0.30, output $1.50, not only for the tokens past 200K. Web search grounding is $2.50 per 1,000 search queries on top of token charges. On a Muse Code subscription, a request over 200K uses the plan's usage faster.

Ember is Standard only, matching Wonder on Model API. Meta publishes no cache-write, Batch, Flex, or Fast price for Ember. Do not invent any.

Peer comparison at list rates

• Under 100K input, Ember, Haiku 5.5, and Luna share the $0.10 / $0.50 sticker. Ember's list-price edge in this band is cache reads at $0.004 vs $0.01.
• From 100K to 200K, Haiku is already on $0.50 / $2.50; Ember and Luna stay on $0.10 / $0.50.
• From 200K to 272K, Ember is on $0.30 / $1.50; Luna is still $0.10 / $0.50; Haiku is $0.50 / $2.50.
• Above 272K, Luna's long-context rate ($0.20 / $0.75) is cheaper than Ember's ($0.30 / $1.50).
• Ember's max output (524,288) is four times Haiku's and Luna's (128,000).
• Ember is 12.5x cheaper on input than Spark Standard ($0.10 vs $1.25) and 40x cheaper than Wonder ($0.10 vs $4.00). That is a list-price ratio, not a cost-per-task.

Default effort is max, and reasoning tokens are billed as output, so a default Ember call can cost more than a default Haiku or Luna call at the same per-token rate. There is no official allowed cost-per-task figure to quantify that. Treat it as a caveat.

Peer list prices: Anthropic pricing for Haiku 5.5 and OpenAI's GPT-6 Luna page. Family rates: Meta pricing and rate limits.

Specs

Reasoning tokens stay private, count toward the output-token budget, and are billed as output. When a request sets no effort, Muse Wonder and Muse Ember use max and Muse Spark uses high. No param count, architecture, training tokens, or knowledge cutoff is published. More detail: reasoning docs, choose a model, and the developer model page.

Safety

Before release, Meta evaluated Ember under its Meta Superintelligence Scaling Framework, including chemical and biological, cybersecurity, and loss-of-control risks, and says Ember remained below the framework's high-risk thresholds. It received the same core safety training as larger Muse models. Meta also evaluated agent behavior (user intent, sensitive information, indirect instructions, asking for human input before high-impact actions). No numeric safety scores were published.

Banner: Meta. Abstract illustration of blue orbital curves, arrows and points looping out from a single center.

FAQ