On November 9, Meta released Muse Ember, its smallest and cheapest model. It targets the work agents repeat all day long: tool calls, programming subtasks, sorting, and extracting data from documents. The price is the same as Haiku and GPT-6 Luna: $0.10 per million input tokens and $0.50 per million output tokens. Cached input costs 60 percent less ($0.004), and above 200 thousand tokens the price triples. It is available in the Meta Model API and in Muse Code; Meta has not released the weights. I tested it through OpenRouter.
Against the competition
Against Haiku 5.5: the same performance for a third of the price. On the Artificial Analysis index both score 36 points. A task costs 14 cents on Ember (about three Czech crowns) and 41 on Haiku. Ember outputs a third of the tokens per task and runs at full effort by default (Haiku runs at medium). Haiku is better at programming, where the benchmark measures whether the code could be merged as-is, and at office work.
Against DeepSeek V4.1 Flash 1017: cheaper, but closed. Ember has three more points and a task costs a quarter as much. On the Vals Index (finance, law, taxes, programming), however, DeepSeek is one point better, and in coding agents the two are equal. DeepSeek has open weights under the MIT license, so you can run it in the EU or on your own hardware, if you have the hardware for it.
Against GPT-6 Luna: eight points better for the same money. Luna costs almost the same per task and is weaker in every test except the science tasks.
Against GPT-6.1 Terra: an important rival that Meta left out. At medium effort Terra scores 38 points for 17 cents per task, so two points more for three cents more. Per token, though, it is ten times more expensive, so on tasks with long input Ember comes out significantly cheaper.
In short, judging by the results, it seems worth deploying Ember for routine data work.
What it’s good for and what to try it on
Extracting data from documents and working with data. In the long-PDF test it beats Haiku and makes little up: on 64 percent of the knowledge questions it prefers to decline to answer. Try it on a set of your own contracts, invoices or statements where you know the correct values, and measure how many fields it extracts exactly.
Process automation and tool calling. In AutomationBench (tasks across business applications) it scores 68 percent, the same as Muse Wonder or GPT-6 Astra. Try it as the executor in an agent where a big model plans the procedure and Ember repeats it a hundred times.
Development helper. In the terminal it is better than Haiku and DeepSeek, and with Muse Code it answers questions about a repository well. Try it as a subagent for searching code, reading logs, writing scripts for cleaning and converting data, and running tests.
What not to use it for. Code that is supposed to go straight into the main branch (there Haiku is eight points better), and research without source material (it knows few facts). Nobody has independently measured its speed yet.
Tests
| At max | Muse Ember | Haiku 5.5 | DeepSeek 1017 | GPT-6 Luna |
|---|---|---|---|---|
| AA index | 36 | 36 | 33 | 28 |
| Cost per task | $0.14 | $0.41 | $0.60 | $0.15 |
| Output tokens per task | 85k | 226k | 139k | 75k |
| Terminal-Bench 4.0 | 41% | 33% | 34% | 13% |
| FrontierCode 2.0 (mergeable code) | 27% | 35% | 28% | 25% |
| Coding Agent Index (with harness) | 46 | 37 | 46 | 31 |
| AutomationBench-AA | 68% | 35% | 67% | 53% |
| GDP.pdf (long documents) | 23% | 21% | 17% | 23% |
| AA-Briefcase (office) | 1521 Elo | 1574 Elo | 1570 Elo | 1337 Elo |
| AA-Omniscience (knowledge) | 7 | 12 | 1 | −8 |
| Vals Index | 60.5% | 58.7% | 61.3% | 52.3% |
| Vals test cost | $0.15 | $3.17 | $0.66 | $0.45 |
| Open weights | no | no | yes, MIT | no |
Sources: Artificial Analysis (index 5.0.1) and Vals AI. The Coding Agent Index measures the model together with its tool: Ember in Muse Code, Haiku in Claude Code, Luna in Codex.
Meta’s table picks its rivals conveniently (both DeepSeek and Terra are missing), and AutomationBench is scored more strictly there, so its numbers can’t be compared with Artificial Analysis. On the main points, though, the independent measurements confirm Meta.
My test: metadata for mobile network specifications
I tried Ember on work I actually do often for the 3GPP Explorer website. The model receives the beginning of an English mobile network specification (about 12 thousand characters) and returns JSON with 22 fields: title, summary, description, category, protocols, interfaces, related specifications. Pure data extraction from a document, exactly what Meta is aiming Ember at.
I took ten specifications and ran each variant once through OpenRouter. Claude Opus scored the answers from 1 to 10 without seeing model names; prices come from actual billing. Reasoning can’t be turned off on Ember, so I went through all six levels. Haiku and GPT-6 Luna ran without reasoning and at higher levels, and the older GPT-5.6 Luna ran the way I use it in production today, that is, without reasoning. The whole thing cost under $2, and the judge ate most of it.
| Variant | Score | Technical accuracy | Cost per 1,000 specs | Median latency |
|---|---|---|---|---|
| Ember, minimal | 8.22 | 7.1 | $0.66 | 2.2 s |
| Ember, high | 8.50 | 7.6 | $1.71 | 5.8 s |
| Ember, max | 8.54 | 7.6 | $2.49 | 7.8 s |
| GPT-6 Luna, reasoning off | 8.32 | 7.5 | $0.64 | 4.1 s |
| GPT-6 Luna, high | 8.30 | 7.6 | $1.13 | 12.2 s |
| GPT-5.6 Luna, reasoning off | 7.65 | 5.4 | $1.36 | 4.8 s |
| Haiku 5.5, reasoning off | 7.70 | 6.0 | $0.94 | 4.2 s |
| Haiku 5.5, high | 8.17 | 6.6 | $2.04 | 12.3 s |
I left the intermediate levels out of the table; they brought nothing interesting.
Against Haiku, Ember wins on both price and quality. At minimum it beat Haiku by half a point, won on eight specifications out of ten, and came out almost a third cheaper. Part of the lead is down to format: Haiku wraps its JSON in markdown and in nine cases out of ten did not respect the requested description length. Without the markdown the difference would be about 0.4 points. The bigger gap is in technical accuracy: according to the judge, Haiku adds protocols and interfaces that don’t belong in the given specification.
With GPT-6 Luna it’s a tie. I did not see the eight-point lead from the Artificial Analysis index on this task. Luna without reasoning scored a tenth of a point better than Ember at minimum, which is pure noise, and the prices are practically the same. Ember responds almost twice as fast. Higher effort didn’t help Luna at all. And the older GPT-5.6 Luna? It is two thirds of a point worse and costs more than double, so it ends here for me and I’m switching to Luna 6.
Beware the default setting. Ember runs at max out of the box. For a thousand specifications you’ll pay $2.50 (about 50 CZK), almost triple Haiku, and get three tenths of a point more than at minimum. I did not measure the one-third price that Artificial Analysis shows: at minimum I save about 30 percent against Haiku. What does higher effort buy? At minimum, the judge faulted Ember four times out of ten for wrongly identifying the 3GPP release in which the specification was created. At max, not once.
Ten specifications is few, and all of them are older (Release 8 and 10), so differences around two tenths of a point may well be noise. I haven’t tried it on texts about 5G and 6G yet.
For this type of work I would pick Ember over Haiku, you just have to turn its reasoning down to the minimum. Against GPT-6 Luna, speed decides, or whatever you already have in place. A harder test is still ahead of it: matching changes between versions of specifications, where some cheap models have already stumbled.
For me it is a useful cheap executor of routine work: bulk agentic steps over documents and data. Anyone who already has Haiku or DeepSeek in place doesn’t have to switch because of Ember. Anyone paying for volume should try it.
Sources: Meta’s announcement, the model page and documentation from November 9, 2026, measurements by Artificial Analysis (index version 5.0.1, Coding Agent Index, AA-Omniscience), Vals AI, and my own test on ten 3GPP specifications. Data valid as of November 9, 2026. The text was written with the assistance of Claude Opus 5.5, that is, a competing model.
