← Back to blog
AiAbout 6 min read

Meta's Muse Ember Has the Same Price List as Claude Haiku and a Third of the Bill

Published Nov 13, 2026
Meta's Muse Ember Has the Same Price List as Claude Haiku and a Third of the Bill

Meta released Muse Ember on November 9, and the first thing to understand about it is that the price list is unremarkable.

Ten cents per million input tokens, fifty cents per million output, four-tenths of a cent for cached input. Those are Claude Haiku 5.5's numbers almost exactly. On a rate card they are the same product.

The claim Meta is actually making is downstream of that. On Artificial Analysis's suite at maximum reasoning effort, Muse Ember scores 36 on the Intelligence Index, level with Haiku 5.5, at an estimated $0.136 per task. Haiku 5.5 scores the same 36 at $0.411. GPT-6 Luna scores 27.8 at $0.152.

Same rate, one third the bill. That gap can only come from one place, and it is the number of tokens each model burns to finish a task. Meta's whole argument is that Ember gets to the same answer with fewer of them.

Where the tokens go

Ember carries a million-token context window and can emit up to 512,000 output tokens in a single call. Both numbers are four times what Claude Haiku 5.5 and GPT-6 Luna allow on output, and the context figure matters more than it looks. A model that has to be re-fed a codebase in chunks pays for the same bytes repeatedly. A model that holds the whole thing pays once.

It also accepts text, images, audio, and video natively, which removes a second kind of duplication: the separate transcription call, the separate OCR pass, the separate vision model for reading a screenshot. Those are small line items individually. Across a hundred thousand agent turns they are not small.

The output cap is the one I keep coming back to. Agent workloads spend most of their tokens producing tool calls, patches, and structured extracts, and a 512K ceiling means the model does not need to compress its work into a series of shorter, context-rebuilding turns. Fewer turns means fewer re-reads of the same conversation.

The benchmarks, with the caveat attached

The spread across benchmarks is wide enough that nobody should quote a single number.

Ember leads on AutomationBench-AA, which covers 657 Zapier-style business workflows, at 37.8% against 20.2% for Luna and 13.9% for Haiku. It leads Terminal-Bench 4.0 at 40.9 against 39.2. It leads the Coding Agent Index, 46 to 37 to 31, and Agents' Last Exam at 61.6. On OSWorld 2.1 it edges Haiku 69.5 to 68.9. PerceptionBench with tools is 80.6 against 78.6 and 62.8, which is the multimodal advantage showing up as a number.

It loses FrontierCode 2.0-AA, 27.2 against Haiku's 34.9. That is a real loss on real code, and it is the kind of loss that matters more than a composite index for anyone shipping software.

There is also a methodology problem worth naming. Meta's evaluation paired Ember with mini-swe-agent, Haiku with Claude Code, and Luna with Codex. Different harnesses, different prompt scaffolding, different retry behavior. Meta says its competitor settings were best effort and might not have been optimally tuned, which is an unusually honest disclaimer and also an admission that the comparison is not clean.

The cost figures come from Artificial Analysis rather than Meta, which helps. They are computed from list prices and measured token mixes, not from vendor projections. But a token mix measured on a benchmark suite is not your token mix, and the per-task cost on your workload will be whatever it is.

The design idea underneath

Ember is the fourth model in the Muse line, after Spark in April, Glimmer in August, and Wonder in October, and it is the first one Meta trained using Wonder Full, its largest model.

That detail explains the shape of the thing. The pitch inside Meta is a division of labor: Wonder plans, Ember executes. A heavier model reads the ambiguous request, decides what the rules are, and hands off; Ember applies those rules to ten thousand cases. That is a distillation story told in product terms, and it is how most large organizations will actually deploy these models, with a smart one at the front and a cheap one in the loop.

The list of launch partners reads like a test of that thesis. Box for document processing, Cline for coding agents, Oracle and Cloudflare for infrastructure, You.com for search. Cloudflare reports up to 30% faster tail latency. You.com reports 2.2 second p95 and 36% to 48% fewer output tokens. Those are the numbers that matter, because they are measured on production traffic rather than a benchmark suite.

The part that should give buyers pause

The weights are private. Ember is available through the Meta Model API, Muse Code, and Google Cloud Vertex AI, and there is no download. For a model positioned as an execution layer inside high-volume pipelines, that is a dependency worth thinking about before it sits under ten thousand daily calls.

The serving surfaces also disagree with each other. Meta says the model handles text, images, audio, and video natively. Google Cloud's preview endpoint, at least at launch, documents text and image only. If you are buying through Vertex AI, test the modalities you actually need rather than assuming the model card transfers.

And the ceiling is real. A 36 on the intelligence index is not a frontier score. Ember is not going to plan a migration or debug a race condition that spans nine services. Meta says so itself, repeatedly, and points anyone with hard problems at Wonder. The mistake to avoid is deploying the cheap model everywhere because the per-task number looks good, then discovering that the tasks it fails are the ones that cost the most to fix.

How to evaluate it in a week

Take a hundred real tasks from your own pipeline, the boring ones: extracting fields from invoices, routing tickets, summarizing call transcripts, patching lint errors. Run them through Ember and through whatever small model you use now. Measure total tokens per completed task and total cost per completed task, not latency and not a benchmark score.

If the token ratio holds anywhere near what Meta claims, the savings are structural rather than promotional, and they compound as volume grows. If it does not, you have spent a week and learned that your prompts were doing the work all along.

Either outcome is more useful than another leaderboard reading. The interesting thing about this release is the unit Meta chose to price. It is charging for the finished task, and that forces everyone else to answer on the same terms.

Related articles