Meta launches Muse Ember, a small multimodal model that undercuts Claude Haiku and GPT-6 Luna on per-task price

November 9, 2026. Meta unveils Muse Ember, a model positioned against Anthropic’s Claude Haiku 5.5 and OpenAI’s GPT-6 Luna. $0.004. That is its price per million cached tokens — a quarter of what its two rivals charge. 69.5% versus 68.9%. That is its partial-credit score on OSWorld 2.1, against Haiku. Why it matters: Ember is not a frontier model — it is the “small tasks” link in an agent chain, and Meta is choosing to fight on cost per task, not on the benchmark.

A small model built for agentic subtasks

Muse Ember joins an already crowded Muse lineup — Muse Wonder Full at the top, then Spark, Code, and Glimmer. Its slot is precise: agentic subtasks. Reading a file, running a test, fixing a targeted bug, summarizing a build output. The overall plan stays with a larger model, Muse Wonder, while Ember executes the repetitive, cheap gestures.

It is a pattern Meta has already used: Ember is the first model trained via distillation from Muse Wonder Full, just as Wonder was used to train the specialized models rolled out inside Meta AI and the Muse personal assistant. Distillation recovers part of a giant model’s behavior in a much smaller — and much cheaper to serve — model.

Ember is multimodal: it accepts text, images, audio, and video as input, with a one-million-token context window and up to 512,000 output tokens. The output window is notable: it allows long responses or long generated files without splitting.

Computer use: driving apps from screenshots

The most concrete part of Ember is its ability to operate desktop applications from screenshots — the computer use family of capabilities. On OSWorld 2.1, a benchmark that measures the ability to complete tasks in a real desktop environment, Ember scores 69.5% partial credit against 68.9% for Haiku, and finishes tasks outright far more often: 33.8% for Ember, versus 27.8% for Haiku and 17% for Luna.

Those numbers deserve caution: on partial credit, Ember and Haiku are essentially tied, separated by a few tenths of a point. The more telling comparison sits elsewhere: Spark 1.3, Meta’s previous model, completed 39.25% of tasks on the same benchmark — above Ember. In other words, Ember is not a raw advance on computer use; it is a better price-performance ratio against its direct rivals.

The real battlefield: the cached-token price

It is on pricing that Meta hits hardest. Ember sells for $0.10 per million input tokens, $0.50 per million output, and — crucially — $0.004 per million cached tokens. The input and output rates line up exactly with GPT-6 Luna and Claude Haiku 5.5 at their standard prices — but those two rivals charge $0.01 per million cached tokens, 2.5 times more.

That detail is decisive for agentic workloads. An agent that re-reads the same context in a loop — a repository’s contents, an error trace, a long file — sends most of its tokens through the cache. On that profile, moving from $0.01 to $0.004 visibly changes the bill. Meta even cites a cost on the order of $0.12 per task on one of its agentic coding benchmarks. The battlefield is shifting: after benchmarks, what now separates small agent models is cost per completed task.

What is still unknown

As with most model launches, several questions remain open. Meta publishes safety and computer use scores but does not specify the model’s exact size or the scale of the distillation data. The OSWorld numbers show a model in the leading pack, not a breakthrough. And the 512,000-token output window, if it holds up in practice, still has to be weighed against real generation costs.

Ember is available in Muse Code and through the Meta Model API. For now it is a closed model — no open-weights announcement, unlike the open-weight strategy Meta pursued with Spark Open and Glimmer. That is a notable choice: Meta appears to reserve openness for its niche models while keeping closed the link that touches the mainstream agent market.

Verdict

If your agentic workloads route most tokens through the cache — re-reading files, logs, long contexts — Muse Ember is worth a trial for its pricing grid alone: at $0.004 per million cached tokens it bills 2.5x less than Haiku 5.5 and Luna on exactly that line item. For production computer use, change nothing on the strength of the scores: Ember, Haiku, and Spark 1.3 sit in the same band, and an internal benchmark on your tasks will weigh more than OSWorld. Keep a frontier model — Muse Wonder, Claude Fable, GPT-6 — for planning, and use Ember only as the executor: that is exactly the role Meta distilled it for.

References

• The New Stack — “Meta launches Muse Ember to take on Claude Haiku and GPT-6 Luna” (Frederic Lardinois, November 9, 2026)
• ETTAYEB — Meta launches Muse Wonder, its first frontier model
• ETTAYEB — Meta launches Muse Code and slashes coding-agent prices tenfold
• ETTAYEB — Anthropic halves the price of Claude Haiku 5.5