Meta has released Muse Ember, which it calls its smallest and most efficient Muse model for the work agents repeat all day: tool calls, coding subtasks, routing, extraction and long document parsing. The research blog is dated 9 November 2026. Mark Zuckerberg posted the same day that Ember “delivers higher intelligence per dollar than others in its class.”
Ember is available as muse-ember-1.0 on Muse Code and the Meta Model API. Meta lists $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens, with a 1M-token context window and up to 512K output tokens. Docs note higher rates above 200K input tokens ($0.30 input / $1.50 output per million). It natively understands text, images, audio and video, and Meta says it is the first model trained using Muse Wonder Full. “Muse Wonder plans, Muse Ember handles the volume.” Meta’s own framing puts Ember under Muse Wonder Full training: Wonder plans the hard work, Ember absorbs the volume of tool calls, coding subtasks, routing, extraction and long document parsing. That is a product hierarchy claim, not an independent finding about which model wins every suite.
Benchmarks Meta published—and where Haiku wins
Every score below is Meta’s. Some evaluations are Meta internal; Meta’s cost chart says Ember’s costs “are estimates at proposed pricing.”
At max reasoning, Meta’s table puts Ember and Claude Haiku 5.5 both at 36 on the Artificial Analysis Index, ahead of GPT-6 Luna at 28. On Terminal-Bench 4.0, Meta reports Ember at 40.9% against Haiku 39.2% and Luna 16.4%. AutomationBench-AA is listed at 37.8% for Ember versus 13.9% for Haiku and 20.2% for Luna.
FrontierCode 2.0-AA is the clear counter-example Meta includes: Haiku leads at 34.9%, with Ember at 27.2% and Luna at 25.2%. That Haiku win belongs in any summary that otherwise leans on Meta’s “intelligence per dollar” framing.
Partner colour on the blog includes Cloudflare CTO Dane Knecht saying Ember ran “up to 30% faster at the tail” than the comparable model Cloudflare tested—again a partner quote, not an independent lab result. Those partner comments sit beside Meta’s tables rather than replacing them. Readers should treat Cloudflare’s “up to 30% faster at the tail” line as colour on one deployment path, not as a substitute for Artificial Analysis or FrontierCode.
Pricing, pairing and safety language
Meta’s portfolio line is explicit: Muse Wonder for the hardest problems, Muse Spark for coding and agentic workflows, and Muse Ember for repeatable coding and agentic tasks at scale. Meta cites about $0.14 per Artificial Analysis task for Ember at its proposed pricing, versus $2.39 for Muse Spark 1.3 on the same comparison table.
On safety, Meta says Ember was evaluated under its Superintelligence Scaling Framework and “remained below the framework’s high-risk thresholds”—Meta’s own evaluation. Related TSN context includes Muse Spaces, the NVIDIA–Meta Muse partnership, and Meta’s Muse bug-bounty payouts.
Small models still inherit the vendor’s scoreboard
Ember’s launch is a pricing and packaging story as much as a capability story. Meta wants a cheap, multimodal workhorse that can sit under Wonder in agent harnesses. The FrontierCode result is useful precisely because Meta printed it: Haiku still wins on that mergeability-style coding suite even while Meta claims Ember matches Haiku’s Artificial Analysis Index at lower estimated cost.
Readers should keep three attributions straight. Benchmarks are Meta’s. Cost-per-task points on the efficiency chart are estimates at proposed pricing. Safety “below high-risk thresholds” is Meta’s framework, not an external audit. The useful next evidence is third-party Artificial Analysis refreshes and harness results that do not use Meta’s preferred setups for every model.
For teams already building on Muse Spark or Wonder, Ember is the volume tier rather than a replacement: route the repeatable tool calls and document passes here, keep the hard planning and long-horizon coding on the larger SKUs, and treat the Artificial Analysis Index tie with Haiku as a Meta-reported headline rather than an independent league table. The MIT-style openness of some rival small models is not the story here: Ember is an API SKU with published rates, a 1M context window and Meta’s own safety-framework language. What matters for buyers is whether the estimated cost-per-task holds once third parties re-run the same harnesses without Meta’s preferred prompts.
Sources
- Meta research blog: Introducing Muse Ember (9 November 2026)
- Meta Model API pricing and rate limits
- Mark Zuckerberg on X, 9 November 2026
Related stories
Meta’s Muse Spaces launch is the nearest prior TSN read on how Meta is packaging Muse for product surfaces beyond a raw model ID.
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
