Meta Muse Ember matches Haiku 5.5 at $0.10 — but costs a third as much per task

Meta has entered the cheapest tier of the model market without undercutting anyone on list price. On November 9, 2026 (US time), Meta Superintelligence Labs released Muse Ember, the smallest and cheapest model in its Muse family and the fourth after Spark (April), Glimmer (August) and Wonder (October). It costs $0.10 per million input tokens and $0.50 per million output tokens — exactly the price of Anthropic's Claude Haiku 5.5 and OpenAI's GPT-6 Luna.

Meta's pitch is not the sticker price but the cost of finishing a task. Artificial Analysis, an independent evaluator, scores Ember 36 on its Intelligence Index at about $0.14 per task. Haiku 5.5 reaches a similar score at about $0.41. Meta's own table also shows where Ember loses: on hard coding problems, Haiku 5.5 is ahead.

1. What shipped

Meta describes Ember as a model for "the work agents repeat all day": tool calls, coding subtasks, routing, extraction and long-document parsing. It is the first model trained using Muse Wonder Full, Meta's largest. The model ID is muse-ember-1.0, with a one-million-token context window and up to 512,000 output tokens. Meta says it takes text, images, audio and video as input. It is available on the Meta Model API and in Muse Code, and is also listed on Google Cloud and Oracle Cloud.

Prices are per million tokens. The three US labs now charge the same headline rate for their small models. The differences sit in the fine print: Ember's higher tier starts at 200K tokens rather than Haiku's 100K, and multiplies the price by three rather than five. For agents that load whole codebases or long documents, that matters more than the list price.

2. Where it wins and where it loses

Meta compared the three models at maximum reasoning settings.

Three caveats apply. First, conditions differ by model: Terminal-Bench was run with mini-swe-agent for Ember, Claude Code for Haiku 5.5 and Codex for Luna, and the Coding Agent Index scores a model together with its harness, so Ember's 46 partly reflects Muse Code. Second, scoring rules move the gap: OSWorld is nearly tied on partial credit but reads 33.8% to 27.8% on a pass-fail basis, and Anthropic's own October figure for Haiku 5.5 on OSWorld 2.1 was 72.4%, not the 68.9% in Meta's table. Third, Ember is weak on factual recall. According to AlphaSignal's summary of Artificial Analysis data, it scored 21.1% accuracy on the AA-Omniscience knowledge test, the lowest among models with a similar index score, and declined to answer 64% of questions — which also kept its hallucination rate, 17.4%, at about half of Haiku's.

3. Why a cheap model for repetitive work exists at all

With identical list prices, a threefold gap in cost per task comes down to token usage. AlphaSignal reports that Ember generated about one-third as many output tokens per index task as Haiku 5.5.

Spark 1.3 scores one point higher than Ember and costs about 17 times as much per task. That gap explains the product. An agent does not answer one question; it reads files, calls tools, checks results and repeats, often hundreds of times. Most of those steps are execution, not judgment, and running each one on a frontier model is unaffordable. The pattern that has emerged is a large model that plans and many small subagents that execute. Meta's version of it: "Muse Wonder plans, Muse Ember handles the volume." AlphaSignal puts Wonder's cost per task at about 50 times Ember's.

The low cached-input rate fits the same logic. In agentic coding, analyses have found that 96–99% of token spending is cache reads — re-reading context the model has already seen — so cache pricing shapes the real bill more than the headline rate.

On speed, Meta quotes early customers: Cloudflare reported tail latency up to 30% lower than a comparable model, and You.com a 95th-percentile latency of 2.2 seconds on coding tasks. These are vendor-selected testimonials. Artificial Analysis lists an output speed of 448 tokens per second. Meta says Ember stayed below its high-risk thresholds for chemical and biological, cybersecurity and loss-of-control risks, and was tested on resisting instructions hidden in web pages and documents; there is no outside verification yet.

4. What remains open

Index scores also shift between versions. These figures use Intelligence Index v5, under which Spark 1.3 scores 37 and Wonder 44; our earlier reports on Muse Spark 1.3 and Muse Wonder cited 61 and 52 from the version current at the time, and the two sets cannot be compared directly. Our reports on Claude Haiku 5.5 pricing and on why AI API prices fell trace how small-model prices reached this level. As planning and execution split between large and small models, the practical choice for developers becomes which lab's large model to pair with which lab's small one.