Meta’s Muse Ember brings lower-cost AI agent and coding work

By Olivia Hart

What does Muse Ember cost?

Meta launched Muse Ember on November 9, 2026, at $0.10 per million input tokens and $0.50 per million output tokens. Meta AI Research introduced the model through Muse Code and the Meta Model API that day, describing it as Meta’s smallest and most efficient option for repeated agent and coding work.

Cached input costs $0.004 per million tokens. Developers can access the model as muse-ember-1.0 through the Responses API, according to Meta.

Meta’s intended tiering positions Ember as the low-cost worker in its model stack, not the premium planner.

What work is Ember built to handle?

Meta targets the repetitive layer of agent activity:

• Tool calls.
• Coding subtasks.
• Routing.
• Extraction.
• Long-document parsing.

Meta says agents spend most of their tokens on these small, structured steps. Ember can retrieve records, call tools in sequence, and update systems across applications, letting it execute a business workflow from start to finish, Meta said.

Its 1M-token context window can accommodate a full codebase or document set in one call, while the model supports outputs of up to 512K tokens, according to Meta.

Meta’s practical claim is simple: use Ember for repeated steps and reserve larger models for the hardest problems.

How does Ember fit Meta’s model family?

Meta’s comparison with Muse Spark 1.3 puts Ember at 36 on the Artificial Analysis Intelligence Index, versus 37 for Spark 1.3. Listed costs per Artificial Analysis task are $0.14 for Ember and $2.39 for Spark. On that basis, Ember costs about 94% less per listed task.

Ember is also the first model trained using Muse Wonder Full. Meta’s intended division puts Wonder on the hardest planning problems, Spark on coding and agentic workflows, and Ember on repeatable versions of those tasks at scale.

Meta says Ember matches Claude Haiku 5.5’s best Intelligence Index result at roughly one-third of its cost per task and beats GPT-6 Luna at every reasoning setting. Meta also says no model on the full Artificial Analysis leaderboard scored higher for less per task. Those cost comparisons are estimates at Meta’s list pricing.

Meta is selling a cost hierarchy, not universal benchmark dominance.

What do the benchmarks show?

At maximum reasoning, Meta reports these results:

• AutomationBench-AA, business workflows: Muse Ember 37.8%; Claude Haiku 5.5 13.9%; GPT-6 Luna 20.2%.
• Agents’ Last Exam, Linux subset: Muse Ember 61.6%; Claude Haiku 5.5 57.0%; GPT-6 Luna 57.1%.
• Terminal-Bench 4.0: Muse Ember 40.9%; Claude Haiku 5.5 39.2%; GPT-6 Luna 16.4%.
• Coding Agent Index-AA: Muse Ember 46; Claude Haiku 5.5 37; GPT-6 Luna 31.
• FrontierCode 2.0-AA: Muse Ember 27.2%; Claude Haiku 5.5 34.9%; GPT-6 Luna 25.2%.
• OSWorld 2.1, partial credit: Muse Ember 69.5%; Claude Haiku 5.5 68.9%; GPT-6 Luna 50.5%.

AutomationBench-AA used strict pass@1 scoring, and Terminal-Bench placed each model in its best published harness. Agents’ Last Exam and OSWorld were Meta internal evaluations.

FrontierCode 2.0-AA evaluates real engineering tasks on mergeability, and Haiku led that benchmark. This result qualifies Meta’s broader coding claim.

Meta’s results show Ember ahead in five of six comparisons, but the FrontierCode loss and two internal evaluations prevent a clean sweep.

Can Ember work across media?

Meta says Ember natively processes text, images, audio, and video. An agent can therefore work from screenshots, charts, scanned documents, voice memos, meeting recordings, or video without calling a separate model. Meta says this extends Ember to desktop computer use and visually detailed documents alongside text-based coding.

Multimodal support broadens the workload; it does not establish equal quality across formats.

What did Meta’s safety tests find?

Meta evaluated Ember under its Superintelligence Scaling Framework for chemical and biological, cybersecurity, and loss-of-control risks. The company reports that Ember stayed below the framework’s high-risk thresholds across the categories assessed and received the same core safety training and testing.

Agent-behavior evaluations covered respect for user intent, sensitive-information handling, resistance to instructions hidden in web pages, documents, or tool outputs, and seeking human input before high-impact actions.

The reported results address stated risks, but they remain Meta’s own assessments.

What have early users reported?

Cloudflare CTO Dane Knecht said: “Ember ran up to 30% faster at the tail than the comparable model his company tested and handled investigations spanning multiple repositories well.”

You.com inference lead Chak Pothina reported “95th-percentile latency of 2.2 seconds,” with Ember completing every coding quest at the lowest latency among the models tested. Pothina also said Ember produced “36–48% fewer output tokens during those tests.”

Those are customer accounts, not an independent production benchmark. The published accounts give no test volumes or shared methodology for either partner result.

For developers, Ember creates a lower-cost execution layer beneath Muse Wonder’s planning and can reduce repetitive calls sent to a more expensive model, Meta said. If Ember’s token savings and latency hold in production, Meta could use it for more routine tool calls and reserve Wonder and Spark for harder work.

That case still rests on Meta’s selected tests and two partners’ early results; the FrontierCode loss shows that cost efficiency does not guarantee every coding win.