要点
- On 9 November 2026, Meta Superintelligence Labs introduced Muse Ember, calling it the company's smallest and most efficient model, aimed at tool calls, coding subtasks, routing, extraction, and long document parsing.
- The model is available as muse-ember-1.0 on Muse Code and the Meta Model API, at $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens, with a 1 million token context window and up to 512,000 output tokens.
- On the Artificial Analysis Intelligence Index, Meta reports Ember at 36, matching Claude Haiku 5.5 and ahead of GPT-6 Luna at 28. Meta puts Ember at about $0.14 per index task, and Haiku's best setting at $0.411.
- Ember leads Meta's table on AutomationBench-AA (37.8 against 13.9 for Haiku and 20.2 for Luna) and on Coding Agent Index-AA (46 against 37 and 31). Haiku leads FrontierCode 2.0-AA, 34.9 against Ember's 27.2.
- Prompts longer than 200,000 tokens cost more: $0.30 per million input tokens and $1.50 per million output tokens. Several headline scores, including Agents' Last Exam and the Personal Superintelligence Index, are Meta's own evaluations.
On 9 November 2026, Meta introduced Muse Ember as its smallest and most efficient model. The post says Ember is built for the work agents repeat all day: calling tools, doing small coding jobs, routing requests, pulling facts out of text, and reading long documents. Meta says Ember delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting, that it matches Haiku's best score at about a third of the cost, and that no model on the full Artificial Analysis leaderboard scores higher for less per task. Those sentences invite a simple misreading, that the cheapest small model is now also the strongest. Meta's own table says otherwise on the hardest coding exam in the set.
[1]What Meta shipped
Muse Ember is sold as a specialist for volume, not as a replacement for Meta's harder models. The post places it next to Muse Wonder, which is meant to plan the hardest problems, and Muse Spark, which is meant for coding and agent workflows. Ember, Meta says, is the first model trained using Muse Wonder Full. It is available today on Muse Code and the Meta Model API, and Meta says Muse Ember, Muse Spark, and Muse Wonder are now available on Google Cloud and Oracle Cloud.
The list price for prompts up to 200,000 tokens is $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens. The context window is 1 million tokens, and the model can write up to 512,000 output tokens. For longer prompts the price rises to $0.30, $0.012 cached, and $1.50. In one chart caption Meta describes Ember's costs as estimates at proposed pricing. A later caption and the pricing section present them as list prices. The distinction matters if you are comparing dollars per task.
Meta also reports early partner comments from Box, Cline, Oracle, Cloudflare, and You.com. Cloudflare's Dane Knecht said Ember ran up to 30% faster at the tail than the comparable model they tested. You.com's Chak Pothina said Ember completed every coding quest they tried with a p95 latency of 2.2 seconds, about 4 times to 13 times faster than leading open-weight models, and with 36% to 48% fewer output tokens. Those figures come from the partners, not from a shared public harness.
[1]What the numbers actually say
The cleanest public comparison is the Artificial Analysis Intelligence Index against cost per task. Meta plots reasoning-effort settings for all three models. Ember appears at $0.116 with a score of 33.9, and at $0.136 with a score of 35.8. Haiku ranges from $0.092 and 24.6 up to $0.411 and 36.2. Luna ranges from $0.073 and 21.4 up to $0.152 and 27.8. The index itself is reported as 36 for Ember, 36 for Haiku, and 28 for Luna. A later comparison puts Ember at $0.14 per index task, against $2.39 for Muse Spark 1.3, which scores 37 on the same index.
On individual tests the picture splits. On AutomationBench-AA, which Artificial Analysis scores as end-to-end business workflows at strict pass@1, Ember's max setting scores 37.8%, against 13.9% for Haiku and 20.2% for Luna. On Coding Agent Index-AA, Ember scores 46, against 37 and 31. On GDP.pdf-AA, professional work over long documents, Ember and Luna both score 22.8%, with Haiku at 20.8%.
Other headline numbers are Meta's. Agents' Last Exam, described as a Linux subset, has Ember at 61.6%, Haiku at 57.0%, and Luna at 57.1%. The Personal Superintelligence Index, an internal composite of eight evaluations, has Ember at 46.8%, against 36.1% and 37.0%. OSWorld 2.1, computer use across desktop apps, is also internal. On partial credit Ember scores 69.5%, Haiku 68.9%, and Luna 50.5%. PerceptionBench with tools is likewise internal: 80.6%, 78.6%, and 62.8%.
Terminal-Bench 4.0 is reported with each model in its best published harness: mini-swe-agent for Ember (40.9%), Claude Code for Haiku (39.2%), and Codex for Luna (16.4%). That is a narrower gap than the marketing sentence suggests, and the harnesses are not the same.
[1]Where Ember does not lead
FrontierCode 2.0-AA asks models to do real engineering tasks that are judged on whether the result could be merged. Haiku scores 34.9%. Ember scores 27.2%. Luna scores 25.2%. If your workload looks like that exam, Haiku is ahead on the number Meta published.
The cost comparison also depends on which Haiku setting you pick. Haiku's best index score, 36.2, sits at $0.411 per task in Meta's chart, about three times Ember's $0.136 at 35.8. A cheaper Haiku setting, $0.248 at 34.3, narrows that gap. Ember still looks cheaper for similar index scores. It does not look cheaper than every Haiku setting at every score.
[1]The fair case, and what is missing
The fair case for Ember is straightforward. Agents spend many tokens on small, structured steps. A model that holds the index score of a strong small rival at a much lower list price, and that leads a public workflow benchmark by a wide margin, is a rational default for those steps. Pairing it with a stronger planner, which is what Meta recommends with Muse Wonder, is the product story.
The missing pieces are the usual ones in a lab announcement. Several of the largest gaps sit on Meta's internal suites. Terminal-Bench compares different harnesses. The first cost chart mentions estimated prices. Independent buyers cannot yet see a full pricing page and an Artificial Analysis run that they themselves commissioned, side by side with Meta's plot. The safety section says Ember stayed below high-risk thresholds in Meta's Superintelligence Scaling Framework for chemical and biological, cybersecurity, and loss-of-control risks. That is Meta's own evaluation of its own model.
[1]Who should care, and what to watch
Teams that route hundreds of subtasks from a larger agent should care, because the buying decision is now about dollars per repeated step, not about winning every public exam. Teams whose work looks like FrontierCode should care for the opposite reason: the cheaper model is not the one that scored highest there.
Over the next six months, watch whether Artificial Analysis, or another third party, republishes the index, the cost-per-task points, and FrontierCode 2.0-AA on the shipped muse-ember-1.0 endpoint at the list prices above. Watch whether production users report that the $0.14-per-task figure holds once prompts grow past 200,000 tokens and the higher price band applies. Watch whether Haiku's lead on FrontierCode remains after both models are run in the same harness.
[1]Close
Ember is a pricing and specialization claim. If independent runs keep the index tie and the FrontierCode gap, then the right default for high-volume agent steps will look different from the right default for mergeable engineering work. If the cost advantage shrinks at production context lengths, or if the internal suites are the only place Ember pulls ahead, then the model will still be a cheap worker, just not the efficiency-frontier story Meta is telling today.
[1]