Meta released Muse Ember on November 9, 2026. It’s the smallest and most efficient model in Meta’s Muse family, built for the work an agent repeats all day: tool calls, coding subtasks, routing, extraction and long document parsing. The API model ID is muse-ember-1.0.
Muse Ember costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 200K tokens. It has a 1M-token context window, can return up to 512K output tokens, and Meta says it handles text, images, audio and video natively.
The list price isn’t what sets Muse Ember apart. Anthropic charges the same $0.10/$0.50 for Claude Haiku 5.5 on prompts up to 100K tokens. Meta’s claim is about cost per finished task: it says Muse Ember matches Haiku’s best Artificial Analysis score at about a third of the cost per task.
Muse Ember pricing and limits
Meta’s announcement lists these rates per million tokens:
| Prompt size | Input | Cached input | Output |
|---|---|---|---|
| Up to 200K tokens | $0.10 | $0.004 | $0.50 |
| Over 200K tokens | $0.30 | $0.012 | $1.50 |
Muse Ember at a glance
| Model ID | muse-ember-1.0 |
|---|---|
| Where | Meta Model API, Muse Code; also Google Cloud and Oracle Cloud |
| Context window | 1M tokens |
| Max output | 512K tokens |
| Inputs | Text, images, audio, video |
| API style | Supports the Responses API |
Meta says Muse Ember is the first model it trained using Muse Wonder Full. In Meta’s lineup, Muse Wonder takes the hardest problems, Muse Spark handles coding and agentic workflows, and Muse Ember covers repeatable coding and agentic tasks at scale.
Muse Ember vs Claude Haiku 5.5 on price
Put the two list prices side by side and they’re almost the same on short prompts. The Haiku 5.5 rates below come from Anthropic’s pricing page.
| Per million tokens | Muse Ember | Claude Haiku 5.5 |
|---|---|---|
| Input, short prompts | $0.10 (up to 200K) | $0.10 (up to 100K) |
| Output, short prompts | $0.50 | $0.50 |
| Cached input / cache hit, short prompts | $0.004 | $0.01 |
| Input, long prompts | $0.30 (over 200K) | $0.50 (over 100K) |
| Output, long prompts | $1.50 | $2.50 |
| Max output | 512K | 128K |
The gaps are in the details. Muse Ember keeps the low rate up to 200K tokens of prompt, while Haiku 5.5 moves to its higher rate after 100K. Ember’s cached input is cheaper. And when a prompt does cross the threshold, Ember’s long-prompt rates are lower than Haiku’s.
For an agent that sends large repository context on every call, that threshold decides which rate most requests pay. Check where your typical prompt size falls before you compare bills.
What Meta’s benchmarks show
Meta published a table at max reasoning effort, comparing Muse Ember with Claude Haiku 5.5 and GPT-6 Luna. Some rows come from Artificial Analysis and some are Meta’s own internal evaluations.
| Benchmark | Source | Muse Ember | Claude Haiku 5.5 | GPT-6 Luna |
|---|---|---|---|---|
| Artificial Analysis Intelligence Index | Artificial Analysis | 36 | 36 | 28 |
| Terminal-Bench 4.0 | Best published harness | 40.9% | 39.2% | 16.4% |
| Coding Agent Index-AA | Artificial Analysis | 46 | 37 | 31 |
| FrontierCode 2.0-AA | Artificial Analysis | 27.2% | 34.9% | 25.2% |
| AutomationBench-AA | Artificial Analysis | 37.8% | 13.9% | 20.2% |
| Agents’ Last Exam | Meta internal, Linux subset | 61.6% | 57.0% | 57.1% |
| OSWorld 2.1 (partial credit) | Meta internal | 69.5% | 68.9% | 50.5% |
Haiku 5.5 still leads on FrontierCode 2.0-AA, the benchmark that judges real engineering tasks on whether the change would be merged. Meta’s table shows Haiku at 34.9% against Ember’s 27.2%. Muse Ember leads on Terminal-Bench 4.0, Coding Agent Index-AA and the agent-workflow tests.
On cost, Meta’s chart puts Muse Ember at about $0.14 per Artificial Analysis task, against $2.39 for Muse Spark 1.3. The chart caption calls Ember’s costs “estimates at proposed pricing”, while the footnote says cost per task uses each model’s list pricing. The scores come from Artificial Analysis runs, but the cost side rests on Meta’s pricing, so treat the per-task dollar figures as estimates.
Where Muse Ember fits in a coding agent
Meta built Muse Ember for the coding subtasks an agent sends off many times in a session: reading files, running tests and making targeted fixes. Meta suggests pairing it with Muse Wonder on the hardest problems, with Muse Wonder planning and Muse Ember handling the volume.
That split is the subagent pattern many coding harnesses support. A larger model plans and reviews, and a cheap model runs the file reads, test runs and small edits, which is where Meta says agents spend most of their tokens. Our Claude Code subagents guide shows how that split works with Anthropic’s models.
Partner quotes in Meta’s post point the same way. Cline founder Saoud Rizwan says Ember suits teams “running it as the subagent in their harness”. Cloudflare CTO Dane Knecht says it ran up to 30% faster at the tail than the comparable model Cloudflare tested. You.com’s inference lead reports a p95 latency of 2.2 seconds in its coding tests, about 4x to 13x faster than leading open-weight models, with 36 to 48% fewer output tokens. These are partner claims from Meta’s announcement, not independent measurements.
Before you route subagent traffic to Muse Ember
- Measure your typical prompt size against the 200K pricing threshold
- Check that your harness can call a Responses API model
- Run your own repo tasks; FrontierCode 2.0-AA favors Haiku 5.5
- Compare cost per finished task, not only per-token price
- Keep a stronger model for planning and review
Safety and data handling
Meta says it evaluated Muse Ember under its Meta Superintelligence Scaling Framework, covering chemical and biological, cybersecurity and loss-of-control risks, and found it below the framework’s high-risk thresholds. Meta also tested how Ember behaves when acting for a user, including whether it resists instructions hidden in web pages, documents or tool outputs and whether it asks for human input before high-impact actions.
A subagent reads untrusted files, web pages and tool output, so prompt injection reaches it first. Meta’s evaluation doesn’t replace sandboxing and permission prompts in your harness.
Is Muse Ember worth testing?
If your agent spends most of its tokens on small, repeated steps, run a trial with Muse Ember. Its list price matches Claude Haiku 5.5 on short prompts, it stays cheap up to 200K tokens of prompt, and Meta’s numbers show it ahead on several agent benchmarks.
If your workload is mostly real code changes that need to get merged, Meta’s own table has Haiku 5.5 ahead on FrontierCode 2.0-AA. Test both on your own repository before switching. For more low-cost options, see our cheapest AI coding API roundup.
