Meta's Muse Ember: The $0.10/M Token AI Model That Takes On Claude Haiku
Meta released Muse Ember on November 9, 2026, and it's the most direct challenge yet to the small-but-capable end of the AI model market. At $0.10 per million input tokens on the Meta Model API, it's positioned to undercut Anthropic's Claude Haiku 5.5 and OpenAI's GPT-6 Luna while matching or beating them on the specific tasks where small models actually get used.
Built for the Boring Parts of Agents
The honest pitch for Muse Ember is that it handles the unglamorous execution layer of multi-step AI workflows. Meta's official announcement describes it as purpose-built for tool calls in sequence, running tests, making targeted code fixes, routing between agents, extracting structured data from documents, and processing long inputs. These are tasks that happen thousands of times in a single agentic pipeline — which is exactly where the per-token cost multiplies fastest.
The model pairs with a larger Muse model (like Muse Wonder) in a two-tier harness: Ember handles the repetitive subtasks, the bigger model handles the reasoning-heavy decisions. Meta says this is the first Muse model it trained using the larger Muse Wonder Full as a teacher, which partly explains its punch-above-its-weight benchmark performance.
1M Token Context, Multimodal Input
Ember ships with a one-million-token context window, which is large even by the standards of frontier models. That's enough to process entire codebases, long legal contracts, or multi-hour meeting transcripts in a single call. It also handles multimodal input: screenshots, charts, scanned documents, voice memos, meeting recordings, and video are all supported. For an agent that needs to read a Jira board, process a Figma export, and summarize a Zoom call without calling three separate specialized models, that breadth matters.
Benchmarks: Better Than Expected
Digital Trends reports that Ember beats Claude Haiku 5.5 on several tests at a fraction of the cost. The New Stack's coverage adds that scores are close to Meta's own Muse Spark 1.3. These are vendor-favorable comparisons — Meta selected which benchmarks to highlight — but the results are consistent across multiple independent analyses: Ember delivers surprisingly strong performance for a model at this price tier.
The model's weight files remain proprietary. Ember is available on the Meta Model API, Muse Code, Google Cloud, and Oracle Cloud under the model ID muse-ember-1.0. This is a closed model despite Meta's history with open weights through the Llama family.
Who This Is For
If you're building anything with agentic loops — workflows where an AI takes dozens or hundreds of sequential steps — Ember is worth benchmarking against whatever you're currently using for the cheap execution tier. At $0.10 per million input tokens, running a million tokens costs a dime. That math changes what's economically feasible at scale.
The model launched about a week before Anthropic's own IPO, which feels like timing. With Claude Haiku 5.5 as a direct named competitor in Meta's own launch materials, the market for "small but capable" AI models is shaping up to be one of the more contested spaces in the near term.
For teams already in Meta's ecosystem — using Muse Code or deployed on Oracle or Google Cloud — the switching cost to try Ember is low. For everyone else, the API is accessible enough that a benchmark run costs under a dollar to set up.