On November 9, 2026, Meta introduced Muse Ember, which it calls its smallest model. It is built for repeatable agent work such as tool calls, coding subtasks, routing, extraction and long-document parsing, and it is priced at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 200K tokens. Meta does not disclose a parameter count, and the page does not state whether the weights are open. The model is offered through Meta's own API, Muse Code and cloud partners, under the API identifier "muse-ember-1.0".
The release is a data point in a broader pattern: as frontier flagships rise in price and capability, the competitive pressure is shifting toward cheap, fast models that handle the high-volume steps inside an agent loop. This post looks at what Meta claims, what the comparison set implies, and where the evidence is limited.
Meta
Anthropic
OpenAIWhat Meta released
Beyond the price, the specification is generous for a small model. Meta lists a 1M-token context window with up to 512K output tokens, multimodal input across text, images, audio and video, and the ability to operate a computer from screenshots. Cached input is priced at $0.004 per million tokens, a 96% discount on the $0.10 input rate. Prompts above 200K tokens move to a higher tier at $0.30 input, $0.012 cached input and $1.50 output.
The tiered long-context price is worth noting. A 1M-token window is a capability, but at the long-prompt rate it costs three times more per input token than at the short-prompt rate, so the headline $0.10 applies only to prompts up to 200K tokens.
Benchmarks: strong on agent tasks, mixed elsewhere
Meta reports results at maximum reasoning against Anthropic's Claude Haiku 5.5 and OpenAI's GPT-6 Luna, the two small models it chose as comparators.
Muse Ember against two small models from rival labs
Scores at maximum reasoning, as reported by Meta on 9 November 2026, per Meta's Muse Ember page.
Self-reported by Meta. Haiku 5.5 leads on FrontierCode 2.0 (34.9% vs 27.2%), a result the chart omits for space. Two of Meta's own evaluations are internal.
The pattern is clear. Ember leads on the agent-oriented evaluations, most dramatically on AutomationBench-AA (37.8% against 13.9% for Haiku 5.5 and 20.2% for GPT-6 Luna), and holds a narrow edge on Terminal-Bench 4.0 (40.9% against 39.2%) and OSWorld 2.1 with partial credit (69.5% against 68.9%). Meta's other reported results include Coding Agent Index-AA at 46 against 37 and 31, and PerceptionBench with tools at 80.6% against 78.6% and 62.8%.
The exceptions matter for interpretation. On FrontierCode 2.0-AA, Haiku 5.5 leads at 34.9% against Ember's 27.2%. On GDP.pdf-AA, Ember and GPT-6 Luna tie at 22.8%. And on the Artificial Analysis Intelligence Index, Ember (36) is level with Haiku 5.5 (36) and ahead of GPT-6 Luna (28). In other words, the claim that survives scrutiny is parity on general intelligence and an advantage on agentic tool-use tasks, not an across-the-board lead.
Two of Meta's reported measures, the Agents' Last Exam Linux subset (61.6% against 57.0% and 57.1%) and the Personal Superintelligence Index (46.8 against 36.1 and 37.0), are Meta-internal. Internal benchmarks are not necessarily biased, but they cannot be checked by outsiders, so they deserve less weight than the third-party-run Artificial Analysis results.
The price claim: cost per task, not just per token
Per-token price comparisons can mislead because models use different numbers of tokens to finish the same task. Meta therefore also reports cost per Artificial Analysis task, about $0.14 for Ember against about $2.39 for its own Muse Spark 1.3.
Muse Ember against Muse Spark 1.3 on price
List price in US dollars per million tokens (prompts up to 200K tokens), per Meta's Muse Ember page.
Meta reports Intelligence Index scores of 36 (Ember) and 37 (Spark 1.3), and about $0.14 against $2.39 per Artificial Analysis task.
Against Muse Spark 1.3 the list-price gap is large: $0.10 against $1.25 per million input tokens, and $0.50 against $4.25 per million output tokens. Yet the Intelligence Index scores are 36 and 37. Taken at face value, that is a model that gives up one index point for roughly a 94% lower cost per task ($0.14 against $2.39). The arithmetic follows from Meta's figures, but the figures themselves are vendor-reported, and cost per task depends on the benchmark mix, not on a team's own workload.
For agent pipelines, the useful question is not which model is smartest. It is which step in the loop needs a smart model at all.
On the economics of small agent models
Why small models matter inside agent loops
An agent that completes a task may make dozens or hundreds of model calls: reading a file, choosing a tool, parsing a result, deciding the next step. Most of those calls do not need frontier reasoning. A common architecture sends the hard planning step to a large model and the routine steps to a small one. If a small model matches a mid-tier one on the routine steps at a fraction of the price, the cost of the whole loop falls sharply, even when the large model is unchanged.
That is the same logic behind Anthropic's guidance on its own new release to use Sonnet or Haiku for volume work and reserve the flagship for the hardest tasks. Meta is competing for the volume tier. Its page does not publish the comparators' list prices, so the comparison here is limited to benchmark scores and Meta's own cost-per-task claim.
Safety and caveats
Meta says it evaluated Muse Ember under its Superintelligence Scaling Framework and found it below the high-risk thresholds in the categories assessed. That is a statement about the categories Meta chose to assess, and the page does not itself publish the underlying results.
Other limits to keep in mind:
- Weights and hosting. Access is through Meta's API and partners, and the page does not say the weights are released, so teams that need self-hosting should not assume it.
- Parameter count unknown. "Smallest model" is a relative claim within Meta's lineup.
- Comparator choice. The two rivals were selected by Meta. Other small models, including open-weight ones, are not in the table.
What to do with this
The practical test is simple. Pick the cheapest model that passes your own evaluation on each step of a workflow, and re-run that evaluation when a new small model ships, because the price-performance frontier at the low end is moving quickly. Using one vendor's model for every step forecloses that option. Multi-model workspaces such as Metir exist for this reason: a team can compare a new small model against its current one on real tasks before changing anything.
Sources:
- Muse Ember - Meta AI Research
- Everything That Happened in AI Today, Tuesday, November 10 - The Neuron
- Claude Fable 5.2 and Claude Mythos 5.2 - Anthropic
Image credits
Header and in-body image: Meta Platforms Headquarters, Menlo Park, California, via Wikimedia Commons, released under CC0 (public domain dedication).