Meta's Muse Ember Does Haiku's Job for a Third of the Cost — Here's the Catch

Nothing here is sponsored — no affiliate links on this page. This breaking article is not the daily buyer-intent post. Facts below come from the sources listed at the end, checked 10 November 2026. We have not tested this hands-on, and nothing here is a hands-on finding.

Meta launched Muse Ember on November 9, 2026 — its smallest and most efficient Muse model, and the clearest sign yet of where the AI industry is heading: not one giant model for everything, but a team of models, with the cheap one doing the bulk of the work. Ember is built for the repetitive jobs AI agents do all day — tool calls, coding subtasks, routing, extraction, long-document parsing. The headline number: $0.10 per million input tokens and $0.50 per million output tokens, with an Artificial Analysis Intelligence Index score of 36 — the same as Claude Haiku 5.5 at maximum reasoning, at about a third of the cost per task.

I have not run Ember yet — nobody outside Meta's early partners had access before yesterday — and this article is built from Meta's announcement and the SiliconANGLE, PYMNTS and Unite.AI coverage of it, checked 10 November 2026. Where a claim is Meta's or Artificial Analysis's, I'll say so, because a launch-day benchmark is a vendor claim, not an independent result.

What actually launched

Muse Ember, model ID muse-ember-1.0, is live now on the Meta Model API and in Muse Code. It is also listed on OpenRouter, and on Google Cloud (through the Gemini Enterprise Agent Platform) and Oracle Cloud. The early partners Meta named — Box, Cline, Oracle, Cloudflare and You.com — tell you who the target customer is: companies running agents at scale, where the model gets called thousands of times a day and the bill is the bottleneck.

The spec sheet is built for volume, not for showpiece reasoning. A 1M-token context window, up to 512K output tokens, multimodal input (text, images, audio and video), and support for the Responses API. Meta's framing is blunt: "Muse Wonder plans, Muse Ember handles the volume." Wonder is the bigger model that figures out what to do; Ember is the one that does the doing, over and over, cheaply.

There is one more detail worth noting: Ember is the first model trained using Muse Wonder Full. Meta is now using its own flagship model to build its smaller ones — a production line, not a one-off launch.

What it costs

The pricing is the whole point of this model, so here it is plainly: $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens. Cached input at four-tenths of a cent per million tokens is effectively free for the repeated prompts agents send — system prompts, retrieved context, the same instructions on every call. That is the pricing shape of a model designed to be called constantly: the per-call cost is so low that the limiting factor becomes your own infrastructure, not the model bill.

For comparison, Meta's own numbers put Ember's cost per task at $0.14 against Claude Haiku 5.5's $0.41 — about a third. And against the other Muse model in the family, Muse Spark 1.3 scores 37 on the same index but costs $2.39 per task, roughly seventeen times Ember's. Spark is the smarter model; Ember is the cheaper one. That is the trade, stated honestly.

How it stacks up — and why to read the numbers carefully

On the Artificial Analysis Intelligence Index, Ember scores 36. Claude Haiku 5.5, at maximum reasoning, also scores 36. Meta says Ember beats GPT-6 Luna at every reasoning setting. And Muse Spark 1.3, the bigger sibling, scores 37 at $2.39 per task.

Read that table the right way round. Ember matching Haiku at a third of the cost per task is a real claim, and if independent testing confirms it, it is a genuine shift in the economics of agent workloads. But every number in that paragraph comes from Meta or from Artificial Analysis — not from independent hands-on testing, and not from us. Benchmarks on launch day are a press release with a chart attached. The honest move is to treat the score as a reason to try Ember on a real workload, not as a settled fact about which model is better.

There is also a structural point worth making. A score of 36 puts Ember in the same tier as Haiku 5.5 on this index — it does not make it a frontier reasoning model. Nobody at Meta is claiming it is. The pitch is volume, not brilliance: a model good enough for the routine work, cheap enough to run constantly.

When Ember is the right pick — and when it is not

Ember makes sense wherever the same kind of call happens over and over: routing a request to the right tool, extracting structured data from documents, running coding subtasks inside a larger agent workflow, parsing long documents, making tool calls in a loop. If you are building agents that do hundreds or thousands of small jobs a day, the per-call price is the feature that matters most — and Ember's price is low enough to change which jobs are worth automating at all.

It is the wrong pick when the task is genuinely hard: novel reasoning, ambiguous problems, work where a wrong answer is expensive. That is what the bigger models are for — Muse Wonder in Meta's own framing, or Opus-class models from other providers. The sensible architecture, and the one Meta is openly selling, is a team: Wonder plans, Ember handles the volume. Using Ember for everything would be the cheap way to get the wrong answer faster; using a flagship model for every routine call is the expensive way to get the right answer slowly.

If you already run agents on Haiku 5.5 or a similar small model, the honest test costs almost nothing: point Ember at your highest-volume, lowest-stakes workload — the one where the output is checked by a human or by a bigger model downstream — and compare the bill and the failure rate over a week. That is the only benchmark that matters for your use case.

The catch, stated plainly

• Every performance claim is Meta's or Artificial Analysis's. The score of 36, the cost-per-task comparison, the claim of beating GPT-6 Luna at every reasoning setting — all vendor-reported or benchmark-provider figures, none independently replicated yet. Treat them as a reason to try the model, not as a settled fact.
• API-only — you cannot download it. Meta has not released the weights, and no parameter count has been disclosed. If you need to run a model locally, self-host it, or fine-tune it on your own hardware, Ember is not an option. This is a rental, not a purchase.
• It is a volume worker, not a thinker. A score of 36 is Haiku-tier, not frontier-tier. The cheap price is the point; do not mistake it for a model that replaces your biggest model on hard problems.
• The ecosystem lock-in is real. Ember lives on the Meta Model API, Muse Code, OpenRouter, Google Cloud and Oracle Cloud. The cached-input price of $0.004 per million tokens is a strong reason to build on Meta's API — and a strong reason to check what happens to your costs if you ever want to leave it.

The bottom line

Muse Ember is the least glamorous kind of launch and possibly the most consequential: a model that is not trying to be the smartest, only the cheapest good-enough option for work that repeats all day. If the Haiku-matching score holds up outside Meta's own testing, the economics of agent workloads shift — not because Ember is brilliant, but because at $0.10 per million input tokens, a whole class of automation that was too expensive to run becomes worth running.

The test is cheap and the verdict is yours to reach in a week: give Ember your highest-volume, lowest-stakes workload and watch the bill and the failure rate. If both stay low, you have found your volume worker.

Would you trust a $0.10-per-million-token model with your highest-volume agent workload — or wait for independent benchmarks first?

Sources: Meta announcement — "Introducing Our Smallest, Most Efficient Model: Muse Ember" (9 November 2026); SiliconANGLE (9 November 2026); PYMNTS (9 November 2026); Unite.AI (9 November 2026); CellCog. All capability, benchmark and pricing claims are Meta's or Artificial Analysis's, stated as such. No hands-on testing was performed for this article. This breaking article is not the daily buyer-intent post, and it contains no affiliate links.