Meta AI has unveiled Muse Ember, which the company describes as its smallest and most efficient model intended for complex reasoning and the repetitive tasks agents perform throughout the day: tool calls, coding subtasks, routing, extraction and parsing of long documents.
Pricing and market position
For shorter prompts, Meta says Muse Ember is priced the same as GPT-6 Luna and Claude Haiku 5.5: $0.10 per million input tokens and $0.50 per million output tokens. All three models, however, have a prompt-length threshold beyond which prices increase, so costs for very long prompts diverge between models.
Reported price tiers for the three models (with model-specific prompt thresholds):
-
GPT-6 Luna
- Prompt size ≤272k tokens: input $0.10/M, cached input $0.01/M, output $0.50/M
- Prompt size >272k tokens: input $0.20/M, cached input $0.02/M, output $0.75/M
-
Muse Ember 1.0
- Prompt size ≤200k tokens: input $0.10/M, cached input $0.004/M, output $0.50/M
- Prompt size >200k tokens: input $0.30/M, cached input $0.012/M, output $1.50/M
-
Claude Haiku 5.5
- Prompt size ≤100k tokens: input $0.10/M, cached input $0.01/M, output $0.50/M
- Prompt size >100k tokens: input $0.50/M, cached input $0.05/M, output $2.50/M
The announcement also contrasts these prices with Claude Fable 5.2, released the same day, which has substantially higher listed rates (about 80× higher on comparable lines at roughly $8/$40). Still, token-based pricing alone can be misleading today because different models may consume very different numbers of reasoning tokens to solve the same task.
Efficiency metrics: Artificial Analysis and the Pareto frontier
Meta published comparisons using metrics such as Artificial Analysis and an "Intelligence Index vs. Cost per Task" chart. According to those visuals, Muse Ember models lie on the Pareto frontier: they represent an efficiency point where improving one dimension (e.g., intelligence) would require sacrificing another (e.g., cost).
Meta highlights that Muse Ember sits on the efficiency frontier of the Artificial Analysis leaderboard, meaning no model scores higher for less cost per task according to that benchmark. The release materials also show other models (for example GPT-6.1 Terra (low)) in positions that the announcement’s author noted as deserving attention.
Capabilities and reasoning levels
Muse Ember offers six reasoning levels: minimal, low, medium, high, xhigh and max. Meta provided examples for different levels; in practice, tests showed that increasing reasoning level did not always produce strictly better or cleaner outputs. For instance, a "high" level attempt failed to return a complete SVG in an illustrative test.
Comparisons in the announcement indicate that some outputs from Muse Ember were similar in quality to results from Muse Wonder, even though Muse Wonder is priced roughly 40× higher than Muse Ember.
Why this matters
The launch indicates growing demand for models that balance low cost with the ability to handle complex reasoning and repeated agent workflows. Understanding stepped pricing and measuring real task-level efficiency — rather than raw per-token costs — will be important for teams selecting models for production, especially when long prompts and repeated tool interactions are involved.
Summary
- Meta AI released Muse Ember 1.0, positioning it as a compact, efficient model for complex and repetitive agent tasks.
- Short-prompt pricing matches GPT-6 Luna and Claude Haiku 5.5, but all three models use tiered pricing that raises costs beyond certain prompt lengths.
- According to Artificial Analysis and intelligence-vs-cost charts published by Meta, Muse Ember sits on the Pareto frontier, offering a strong efficiency trade-off between performance and per-task cost.
- The model supports six reasoning levels and up to one million input tokens, though higher reasoning levels do not always map neatly to visibly better outputs.
Muse Ember provides a cost-competitive alternative for developers and service providers focused on long-running agent workflows and complex reasoning tasks.
