The Pulse
Meta Launches Muse Ember at $0.10 per Million Input Tokens
Muse Ember, Meta's smallest Muse model, targets agent subtasks at Haiku 5.5's headline prices. Artificial Analysis scores it 36 at about $0.14 per task; several of Meta's comparisons use different harnesses.
AI.info Team · Written with AI · fact-checked ·
Meta has added a smaller model to its Muse line. Muse Ember, released on November 9, is the company's smallest and cheapest model, aimed at the repetitive steps an AI agent takes over and over: calling tools, handling coding subtasks, routing requests, pulling data out of documents and parsing long files. Meta lists it at $0.10 per million input tokens and $0.50 per million output tokens, the same headline rates Anthropic charges for Claude Haiku 5.5 on shorter prompts.
The pitch is cost per unit of capability rather than raw capability. Meta says Ember gives more intelligence per dollar than Haiku 5.5 or OpenAI's GPT-6 Luna at every reasoning setting it charted, and that it reaches roughly the score of Haiku's strongest setting for about one third of the cost.
What Meta Is Selling, and at What Price
The model is available through the Meta Model API and Muse Code, Meta's coding agent, under the identifier muse-ember-1.0. It accepts up to 1 million tokens of context and can return up to 512,000 output tokens, and Meta says it works with the Responses API.
The $0.10 and $0.50 rates are not the whole price list. Meta's Muse Code pricing table splits Ember into two tiers by context length:
- Requests under 200,000 tokens of context: $0.10 per million input tokens, $0.004 per million cached input tokens, $0.50 per million output tokens.
- Requests over 200,000 tokens of context: $0.30 input, $0.012 cached input and $1.50 output, per million tokens.
Anthropic's Haiku 5.5 uses a similar two-step scheme with a lower threshold: $0.10 and $0.50 for prompts up to 100,000 tokens, and $0.50 and $2.50 above that, with cache reads at $0.01 in the lower tier, according to Anthropic's pricing table. On long prompts, Ember's listed rates are lower than Haiku's; on short prompts the input and output rates match and Ember's cached input is cheaper.
Within Meta's own catalog, the gap is wider. Muse Spark 1.3, which Meta offers for coding and agent work, is listed at $1.25 per million input tokens and $4.25 per million output tokens for the version whose data is not used to improve Meta's products.
The Benchmark Case, and Its Footnotes
Benchmarking firm Artificial Analysis gives Ember at maximum reasoning effort a score of 36 on its Intelligence Index, against a median of 24 for models in a comparable price tier, and puts its cost at about $0.14 per index task. On Meta's comparison table, Haiku 5.5 also scores 36 on the same index and GPT-6 Luna scores 28. Muse Spark 1.3 scores 37, one point higher, but costs $2.39 per task, Meta says, citing Artificial Analysis figures.
Meta's own chart plots each reasoning setting separately. Ember's maximum setting scores 35.8 at $0.136 per task, while Haiku 5.5 reaches 36.2 at $0.411. That is the basis for the "about a third of the cost" claim. Meta's chart caption describes Ember's costs as estimates at proposed pricing, while the footnote says cost per task uses list pricing.
Low token prices do not mean few tokens. Artificial Analysis says Ember produced 115 million output tokens while running the index, which it calls somewhat verbose against a median of 90 million for its tier.
Meta's broader benchmark table, all at maximum reasoning, shows Ember ahead in most rows and behind in one:
- AutomationBench-AA, Zapier's private set of 657 business-workflow tasks run by Artificial Analysis: Ember 37.8%, GPT-6 Luna 20.2%, Haiku 5.5 13.9%.
- Terminal-Bench 4.0: Ember 40.9%, Haiku 5.5 39.2%, Luna 16.4%.
- OSWorld 2.1, computer use on a desktop, partial credit: Ember 69.5%, Haiku 5.5 68.9%, Luna 50.5%. On the strict measure of fully completed tasks the figures are 33.8%, 27.8% and 17.0%.
- FrontierCode 2.0-AA, real repository tasks judged on whether a maintainer would merge them: Haiku 5.5 leads at 34.9%, with Ember at 27.2% and Luna at 25.2%.
The methodology report Meta published alongside the launch explains why some of these numbers are not like-for-like. On Terminal-Bench 4.0, Ember's result is Artificial Analysis's run in the mini-swe-agent harness, Haiku's is Anthropic's self-reported figure from Claude Code, and Luna's is the Codex result on the public leaderboard. On Agents' Last Exam, Meta ran the Linux task split and left out five of the benchmark's 105 tasks for every model. OSWorld 2.1 was run in an internal framework rather than the official runner. Meta also calls its internally generated competitor runs best-effort, noting that prompts, tools and runtime may not be tuned for rival models. Several rows, including Agents' Last Exam, OSWorld and PerceptionBench, are Meta's own evaluations rather than third-party results.
Built From Muse Wonder Full
When Meta introduced Muse Wonder and Muse Wonder Full, it said it would draw on the pair to build specialized models it could make available to everyone. Ember is the first model trained using Muse Wonder Full, Meta says. The company now describes a three-part lineup: Wonder for the hardest problems, Spark for coding and agentic workflows, and Ember for repeatable tasks at scale. It suggests pairing them, with Wonder planning and Ember handling volume.
Ember is not an open-weight release: Artificial Analysis lists it as proprietary and says Meta has not disclosed its size or parameter count.
Meta says Ember natively understands text, images, audio and video. Artificial Analysis's listing records text and image input with text output.
What Early Testers Reported
Meta's announcement carries comments from Box, Cline, Oracle, Cloudflare and You.com, which tested the model before release. Cloudflare's chief technology officer pointed to latency:
What stood out in our testing was speed. Muse Ember ran up to 30% faster at the tail than the comparable model we tested, and it handled broad, multi-repo investigation tasks well.
Dane Knecht, CTO, Cloudflare
You.com's inference lead, Chak Pothina, said Ember finished every coding task in its tests with a p95 latency of 2.2 seconds, which he described as “about 4x to 13x faster than leading open-weight models,” with 36–48% fewer output tokens. These are partner results published by Meta, not independent tests.
Safety Testing
Meta says it evaluated Ember under its Meta Superintelligence Scaling Framework, covering chemical and biological, cybersecurity and loss-of-control risks, and found it stayed below the framework's high-risk thresholds in the categories assessed. The company says the small model received the same core safety training as its larger ones, and that it tested behavior when acting for users, including resistance to instructions hidden in web pages, documents or tool outputs and asking for human input before high-impact actions.
Meta closes by saying it has more efficient models in training. Artificial Analysis lists one API provider for Ember, while Meta says Muse Ember, Muse Spark and Muse Wonder are now available on Google Cloud and Oracle Cloud.