What Meta shipped
On 9 November 2026 Meta released Muse Ember, the smallest and cheapest model in its Muse family. Meta describes it as a model for the high-volume work inside agent systems: tool calls, routing, extraction, coding subtasks and parsing long documents. Its API name is muse-ember-1.0. Like the rest of the Muse line, it's closed-weight, so you can only reach it through hosted APIs.
The model docs are clear about how Meta wants it used. They tell developers to pair it with Muse Spark or Muse Wonder, with the bigger model doing the planning and Ember doing the bulk of the work. That's a familiar pattern in agent design. One capable model splits a job into steps, and a cheaper model runs those steps many times over. Ember is Meta's entry in that slot, and Meta says it's also the first model trained using Muse Wonder Full.
Specs and how you call it
Meta's announcement gives Ember a 1 million token context window and up to 512K output tokens. Vals AI lists the output ceiling as 524K, which is the same figure counted in binary thousands. The developer docs list text, image, video, audio and PDF as inputs, with text as the only output. Meta's blog also says Ember can operate a computer from screenshots and handle material like scanned documents and voice memos.
There's one mismatch to know about. Artificial Analysis lists Ember as taking text and image input only. That may just mean its tests didn't cover audio and video, but if you're building on those inputs, test them yourself before you rely on them.
Ember is a reasoning model with adjustable effort. Artificial Analysis tested two settings, "Max" and "Xhigh". The model docs list Responses, Chat Completions and Messages as supported endpoints. On Meta's Model API, Ember runs in Standard mode only. Meta hasn't offered the cheaper Contributor mode for it, which on Muse Spark trades lower prices for permission to train on your prompts and outputs.
The pricing, and where the cliff sits
For requests with up to 200K input tokens, Ember costs:
- $0.10 per million input tokens
- $0.004 per million cached input tokens
- $0.50 per million output tokens
Above 200K input tokens, every rate triples: $0.30 for input, $0.012 for cached input and $1.50 for output.
The detail that matters most is how the threshold is applied. Meta's pricing page says the tier depends on the total input token count of the request, and cached tokens count toward that total. Once a request goes past 200K, the higher rate covers the whole request, including its output, and not only the tokens above the line. Requests that continue a stored conversation through previous_response_id also count that stored history as input. So a long agent session can cross the line without anyone noticing, and from then on every turn costs three times as much.
Muse Spark doesn't work this way. It charges a flat $1.25 input, $0.15 cached input and $4.25 output per million tokens, however full its context window gets. Even at its long-context rate, Ember is still far cheaper than Spark per token, so the cliff mostly matters when you're comparing Ember with itself or with rival small models.
Rate limits are generous for a new model. Meta lists 3,000 requests per minute and 4 million tokens per minute for muse-ember-1.0, the same as Standard-mode Spark. Limits apply per team and per model, so adding API keys won't raise them. For 429 errors Meta recommends exponential backoff with jitter and respecting any Retry-After header.
How it compares with Muse Spark
On the Artificial Analysis Intelligence Index, Ember at Max effort scores 36 and Muse Spark 1.3 scores 37. The cost gap is much bigger than that one-point difference. Meta puts Ember's cost per index task at about $0.14 against $2.39 for Spark. Artificial Analysis shows a similar figure for Ember, $0.14 at Max and $0.12 at Xhigh, where it scores 34.
Speed looks good as well. Artificial Analysis measured about 446 output tokens per second, with time to first answer token between roughly 4 and 5 seconds depending on effort setting. Meta still points to Spark for open-ended coding and agentic work, and to Ember for repeatable tasks run at volume.
Benchmarks: whose numbers are these?
Meta's launch table compares Ember with Claude Haiku 5.5 and GPT-6 Luna at maximum reasoning. Ember leads on most rows. It scores 37.8% on AutomationBench-AA against 13.9% for Haiku, 40.9% on Terminal-Bench 4.0 against 39.2%, and 69.5% on OSWorld 2.1. It trails Haiku on FrontierCode 2.0-AA, 27.2% to 34.9%, and ties GPT-6 Luna on GDP.pdf-AA at 22.8%.
Read the footnotes before you quote any of these. Many of the results come from Meta's own evaluations. The Agents' Last Exam score covers only the Linux subset. Terminal-Bench uses each model's best published harness, so the setups differ from model to model. The OSWorld figures use partial credit, and Meta's full table also shows stricter binary scores.
The independent results are more mixed. Artificial Analysis ranks Ember Max second of 36 models in its class on its Intelligence Index. Vals AI paints a middling picture:
- Vals Index: 60.48%, around 12th or 13th of 36
- Vals Legal Index: 75.69%, 4th
- Vals Finance Index: 60.91%, around 11th
- Terminal-Bench 4.0: 39.39%, 16th of 63, slightly below Meta's own 40.9%
- Vibe Code Bench v1.1: 83.70%, 23rd of 127
Vals says Ember had the lowest cost per test of any model on its index, at $0.153. The ranks on the Vals page don't quite agree with each other (the table and the update notes differ by one place on several benchmarks), so we've given approximate positions.
Partner claims from Cloudflare and You.com about tail latency and fewer output tokens come from early testing and haven't been independently verified, so we'd treat them as anecdotes.
Safety notes
Meta says it tested Ember under its Superintelligence Scaling Framework for chemical and biological, cyber and loss-of-control risks, and that it stayed below the high-risk thresholds in every category assessed. Meta also says it checked how well the model resists hidden instructions in web pages and tool outputs, and whether it asks a person before taking high-impact actions. Those are useful checks for a model that's meant to run unsupervised inside agent loops. They're still Meta's own assessments.
Access from New Zealand and Australia
You can reach Ember through Meta's Model API and Muse Code, and through Google Cloud and Oracle Cloud. OpenRouter lists two providers, Meta and Google Vertex, both at $0.10 and $0.50, and says EU in-region routing isn't available.
Meta's Geographic Use Policy, last updated 12 October 2026, lists restricted territories such as China, Russia and Iran. New Zealand and Australia aren't on that list. The policy also doesn't explicitly confirm them as supported, and Meta says it can change access at any time. Check access from your own account before you commit a production workload.
What builders can do with it
- Use it as the worker in planner and worker setups. Keep Spark or another strong model for planning, and send the repetitive calls (classification, extraction, tool execution) to Ember.
- Watch the 200K line. Log input token counts for every request, cached tokens included. Trim or summarise conversation history before it reaches the threshold, and be careful with
previous_response_idchains. - Run your own evals. The benchmarks don't all agree, so compare Ember with Haiku 5.5 or your current small model on a sample of your real tasks before you switch.
- Test the effort settings. Xhigh was cheaper and quicker per task than Max in Artificial Analysis testing, for a two-point drop in score. For many subtasks that's a good trade.
- Weigh the closed weights. You can't self-host Ember, so data residency and vendor lock-in still apply.
Sources
- Muse Ember announcement, Meta
- Pricing and rate limits, Meta Model API docs
- Models, Meta Model API docs
- Geographic Use Policy, Meta Model API
- Muse Ember, Artificial Analysis
- Muse Ember, Vals AI
- Meta: Muse Ember, OpenRouter
Build with people working on the same things
Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.