On 9 November 2026 Meta released Muse Ember, which it calls its smallest and most efficient model. Meta says Ember matches the best Artificial Analysis Intelligence Index score of Claude Haiku 5.5 at about a third of the cost per task. A third of the cost is a cut of about two-thirds, and that comparison is the index, not a coding-only bill.
Meta launches its smallest AI model. The lightweight Muse Ember system cuts operating costs by two-thirds for coding tasks. The coding tests in the same launch do not all point the same way: Ember leads some of them, and Haiku leads the one that asks whether a code change is ready to merge. Figures are as of 11 November 2026. Bdeb Technology has not run the model.
Meta launches its smallest AI model. The lightweight Muse Ember system cuts operating costs by two-thirds for coding tasks.
The research account is Meta’s 9 November launch note. The developer account, with prices for long prompts and two worked examples, is the same-day build note. Meta says Ember is for the steps an agent repeats all day: tool calls, coding subtasks, routing, extraction, and long documents. For harder planning it says to pair Ember with Muse Wonder or Muse Spark.
Where the two-thirds comes from
Meta’s sentence is that Ember delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting, and that it matches Haiku’s best score at about a third of the cost. In the summary table, both Ember and Haiku are scored 36 on the Artificial Analysis Intelligence Index, and Luna is scored 28. The cost line in that table is about $0.14 per index task for Ember. The plotted points on the same page are slightly different: Haiku’s highest plotted score is 36.2 at $0.411 a task, and Ember’s higher plotted point is 35.8 at $0.136. Dividing 0.136 by 0.411 is about one third, so the cut is about two-thirds. The developer note, written the same day, says the max-effort cost is about $0.13 a task. Meta has published three nearby figures, $0.13, $0.136, and about $0.14, for the same idea.
That is not a measurement of a coding invoice. Meta does not publish a separate “coding tasks cost two-thirds less” number. The token sticker is also not cheaper than the class it is compared with. Ember’s list price, up to 200,000 input tokens, is $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens. Above 200,000 input tokens those rates become $0.30, $0.012, and $1.50. The lower cost per index task is the measured result, from how many tokens the run used and from the cache price, not from a lower headline rate than Haiku.
What the coding tests show
All of these are at maximum reasoning, from Meta’s launch table. Terminal-Bench 4.0 uses each model’s best published harness, which Meta names as mini-swe-agent for Ember, Claude Code for Haiku, and Codex for Luna. They are not the same harness.
| Test | Muse Ember | Claude Haiku 5.5 | GPT-6 Luna |
|---|---|---|---|
| Artificial Analysis Intelligence Index | 36 | 36 | 28 |
| Cost per index task, summary table | About $0.14 | Haiku’s best plotted point is $0.411 | Not in the summary cost cell |
| Terminal-Bench 4.0 | 40.9 percent | 39.2 percent | 16.4 percent |
| Coding Agent Index | 46 | 37 | 31 |
| FrontierCode 2.0, judged on whether a change is ready to merge | 27.2 percent | 34.9 percent | 25.2 percent |
| AutomationBench, strict pass at 1 | 37.8 percent | 13.9 percent | 20.2 percent |
Ember is ahead on the terminal test and on the coding-agent index. Haiku is ahead on FrontierCode, the test of merge-ready engineering changes. A team that sends Ember the small, checked steps and keeps a larger model for the plan is the split Meta describes. A team that treats the two-thirds figure as the saving on every coding job is reading past what Meta published.
The same table includes tests Meta marks as its own. Agents’ Last Exam, a Linux subset, is 61.6 percent for Ember, 57.0 for Haiku, and 57.1 for Luna. The Personal Superintelligence Index, an internal composite, is 46.8, 36.1, and 37.0. OSWorld 2.1, scored with partial credit, is 69.5, 68.9, and 50.5. Meta also gives a stricter all-or-nothing OSWorld score of 33.8 for Ember. PerceptionBench with tools, also internal, is 80.6, 78.6, and 62.8. GDP.pdf, from Artificial Analysis, is 22.8 percent for Ember, 20.8 for Haiku, and 22.8 for Luna.
Price, context, and where it runs
The model id is muse-ember-1.0. The context window is 1 million tokens, and output can run to 512,000 tokens. Meta says it reads text, images, audio, and video. It is on Muse Code and the Meta Model API, and Meta says Muse Ember, Muse Spark, and Muse Wonder are on Google Cloud and Oracle Cloud. Against Muse Spark 1.3, Meta’s comparison table puts Ember at about $0.14 per index task and Spark at $2.39, with Spark half a point higher on the index, 37 against 36. Spark is a different class of model, at $1.25 per million input tokens and $4.25 per million output tokens in that table.
The build note includes two Meta-run examples, and it warns that results vary. A Python-to-Rust port of 16 Pygments lexers, with Muse Wonder planning and 16 Ember workers, matched Python on all 383 test inputs and on 62 of 62 held-out files. Meta says the Rust commands were 3.0 to 7.3 times faster than Python on about 1 MB of input, that Ember handled 158 million input tokens against Wonder’s 4.4 million, and that the run cost about $3.96, of which $1.84 was Ember and $2.12 was Wonder. A sort of 10,000 photos is given two totals in that same note: $1.59 with $0.50 of it Ember, and later $1.33 with $0.52 of it Ember. Meta also says the same photo requests at Wonder’s price would have been about $20. The two photo totals conflict, and Meta does not say which one replaces the other.
Cline’s founder, quoted by Meta, says teams are cutting costs by running Ember as the subagent. He does not give a percentage. Cloudflare’s CTO says Ember ran up to 30 percent faster at the tail than the comparable model Cloudflare tested. Those are partner comments published by Meta, not an independent bill.
On safety, Meta says Ember stayed below the high-risk thresholds in its Meta Superintelligence Scaling Framework for chemical and biological risk, cybersecurity, and loss of control, and that the small model got the same core safety training. For where a small coding model sits next to the tools already covered on this site, see how to choose an AI coding tool, the Fable 5.2 and Mythos 5.2 notes, and DeepSeek API pricing.
Sources
- Meta, Introducing Muse Ember, 9 November 2026. The index scores, the cost points, the coding table, the long-prompt prices, the model id, and the safety paragraph.
- Meta, Start building with Muse Ember, 9 November 2026. The $0.13 figure, the lexer port, and the two photo-job totals.
- Artificial Analysis, Muse Ember and Claude Haiku 5.5. The independent index Meta cites. Its max-effort row also shows 36 and about $0.14 against about $0.41.
Meta launches its smallest AI model. The lightweight Muse Ember system cuts operating costs by two-thirds for coding tasks.
Meta released Muse Ember on 9 November 2026 and calls it the company’s smallest model. It says Ember matches Claude Haiku 5.5’s best Artificial Analysis Intelligence Index score at about a third of the cost per task, which is a cut of about two-thirds. That figure is the index, not a coding invoice. On coding tests Ember leads Terminal-Bench 4.0 and the Coding Agent Index, and Haiku leads FrontierCode 2.0.
What does Muse Ember cost?
Up to 200,000 input tokens, Meta’s price is $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens. Above that, the rates are $0.30, $0.012, and $1.50. The model id is muse-ember-1.0.
Where can developers use it?
Meta says Ember is on Muse Code and the Meta Model API, and that Ember, Muse Spark, and Muse Wonder are on Google Cloud and Oracle Cloud. The context window is 1 million tokens, with up to 512,000 tokens of output.