Meta's Muse Ember Beats GPT-6 Luna at 13x the Speed
Meta's Muse Ember now leads efficient models on Arena's Agent Arena , ahead of GPT-6 Luna and Claude Haiku 5.5. Ember finishes tasks in 27 seconds versus 5.8 minutes for Luna and 1.9 minutes for Haiku. It costs 1.9¢ per task versus about 7¢ for rivals and has +2.3% Net Improvement.

Takeaways
- Meta's Muse Ember now leads efficient models on Arena's Agent Arena , ahead of GPT-6 Luna and Claude Haiku 5.5.
- Ember finishes tasks in 27 seconds versus 5.8 minutes for Luna and 1.9 minutes for Haiku.
- It costs 1.9¢ per task versus about 7¢ for rivals and has +2.3% Net Improvement.
- Its steering score (fewer user corrections) is +11.2% overall and +12.0% on document tasks.
- Pricing is $0.10/$0.50 per million tokens with 1M context, tripling above 200K input tokens.
- Caveats: modest overall rank, trails Haiku on FrontierCode, and it's closed and API-only.
Meta’s Muse Ember ranks first in the efficient tier of Arena’s Agent Arena leaderboard , ahead of OpenAI’s GPT-6 Luna and Anthropic’s Claude Haiku 5.5 on task quality, speed, and cost per task. Arena draws its scores from real agent sessions, giving developers evidence about how these models perform under user-driven workloads.
At the listed reasoning settings, Ember completes a typical task in 27 seconds. Luna takes 5.8 minutes, and Haiku takes 1.9 minutes. Ember costs 1.9¢ per task, compared with about 7¢ for each rival. Those figures make Ember about 13 times faster than Luna and four times faster than Haiku while costing less than one-third as much.
A worker model taught by Wonder
Meta Superintelligence Lab describes Muse Ember as its lightest and most cost-efficient model. The company designed it for high-volume agent work such as tool calls, coding subtasks, request routing, data extraction, and long-document parsing.
Meta positions Ember as the worker in a two-model architecture: “Muse Wonder plans, Muse Ember handles the volume.” Ember is also the first model Meta trained using Muse Wonder Full, its largest model, allowing the larger system to help teach the smaller one.
What developers get
- Inputs and context: Accepts text, images, video, audio, and PDF documents, returns text, and supports a 1 million-token context window.
- Standard pricing: $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens.
- Access: Available through the Meta Model API, Muse Code, and services that resell the API. Meta does not offer downloadable model weights.
- Model ID:
muse-ember-1.0
Arena scores live agent work
Agent Arena evaluates real Agent Mode interactions using causal tracing, a statistical method that estimates how outcomes change when one model replaces another. Its headline metric, Net Improvement, combines several behavioral signals into a score relative to the average model on the platform. As the average model improves, a model’s Net Improvement score can decline even when its underlying performance stays constant.
Arena’s contributing signals cover confirmed task success, user praise and complaints, mid-task corrections, Bash command reliability, and hallucinated tool use. Together, they capture whether an agent completes work correctly and how much intervention it requires.
| Metric | Muse Ember (Max) | GPT-6 Luna (Max) | Claude Haiku 5.5 (High) |
|---|---|---|---|
| Time per task | 27 seconds | 5.8 minutes | 1.9 minutes |
| Net Improvement | +2.3% | +1.4% | +0.4% |
| Cost per task | 1.9¢ | 7.0¢ | 6.9¢ |
| Steering, all work | +11.2% | +3.9% | +3.8% |
| Steering, document tasks | +12.0% | +0.4% | -0.1% |
Max and High are vendor-defined reasoning settings, so they do not guarantee equal inference budgets. The comparison reflects each model’s published configuration on Arena.
Steering creates the widest gap
Ember’s largest advantage appears in steering, Arena’s measure of how often users must correct a model during a task. A higher score indicates fewer interventions relative to the platform average. Ember records +11.2% across Arena’s all-work category and +11.6% on software and IT tasks, roughly three times the scores of Luna and Haiku. On document work, Ember reaches +12.0%, while Luna records +0.4% and Haiku records -0.1%.
Arena reports successful Ember sessions across several kinds of work:
- Writing search-optimized product listings
- Reviewing pull requests
- Preparing client reports about tax law
- Checking suspicious files for malware
Seventy percent of successful Ember tasks finish in less than two minutes and cost under 5¢, according to Arena. In a separate evaluation, penetration-testing company Arnica reported that Ember found 31% more known vulnerabilities, completed tests 54% faster, and consumed 71% fewer reasoning tokens than its frontier-model baseline. Arnica used its own workload and baseline, which limits direct comparison with Arena’s results.
Output tokens set the bill
Ember, Haiku, and Luna list the same base rates for uncached input and output tokens. Ember’s cost advantage comes from generating less text. Artificial Analysis measured about 84,000 Ember output tokens per Intelligence Index task, compared with roughly 226,000 for Haiku at its maximum setting. Ember’s cached-input rate, which accounts for most spending in its agent evaluations, is also 60% lower than the corresponding rates for Haiku and Luna.
The Artificial Analysis Intelligence Index combines several benchmark results into one score. Ember and Claude Haiku 5.5 each receive a rounded score of 36, while GPT-6 Luna scores 28. Estimated cost per task is about $0.14 for Ember, $0.41 for Haiku, and $0.15 for Luna.
A one-line model swap
Ember accepts the same API request schema as Muse Spark, so many existing integrations require only a model ID change. This minimal Python example calls the Responses endpoint shown in the official docs :
The lead has limits
Ember’s +2.3% Net Improvement score leads the efficient tier but remains far below the overall leaderboard’s strongest result. Claude Fable 5.2 recorded 28.61% ± 3.72% in the cited Nov. 13 snapshot. Error bars for smaller models can also exceed the gaps between their scores, so the current +2.3% versus +1.4% result does not establish a decisive quality advantage.
- Results vary by workload. Ember beats Claude Haiku 5.5 on Terminal-Bench 4.0, scoring 40.9% versus 32.8%, but trails it on FrontierCode 2.0, scoring 27.2% versus 34.9%. The rounded Intelligence Index tie also conceals a small difference: Meta’s chart lists 35.8 for Ember and 36.2 for Haiku.
- Long contexts trigger higher rates. Requests containing more than 200,000 input tokens cost $0.30 per million input tokens and $1.50 per million output tokens, three times the standard rates.
- Early API failures occurred. Arnica recorded four failed Ember calls among 24,416 requests, an observed failure rate of about 0.016%. Its comparison models recorded none.
- Some tasks favor larger models. Meta recommends Muse Spark or Muse Wonder for difficult single-shot coding problems, knowledge-heavy questions, and visual coding.
- Deployment remains hosted. Teams that require local or on-premises inference cannot deploy Ember’s weights and must assess Meta’s API against their data-handling requirements.
Route by workload, then measure
For agent systems, total task cost depends on token volume, cache pricing, retries, and correction rates alongside the advertised per-token rate. Meta’s own lineup illustrates the difference: Ember scores 36 on the Intelligence Index and costs about $0.14 per task, while Muse Spark 1.3 scores 37 and costs about $2.39.
A production evaluation should route a representative share of worker-agent traffic to Ember while keeping prompts, tools, and retry policies consistent. The comparison should track:
- Task success and the number of user corrections
- Median and tail latency, including p95 response time
- Input, cached-input, reasoning, and output token use
- API failures, retries, and incomplete tool calls
- Total cost per successful task
- Performance on requests above the 200,000-token pricing threshold
Agent Arena’s ranking will change as new sessions arrive, uncertainty narrows, and competitors release updated models. Ember’s current results support its use for high-volume routing, extraction, document processing, and coding subtasks, with production telemetry determining whether its speed and cost advantages hold for a specific workload.
Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.
Read original report

