If you run high-volume agent subtasks on Claude Haiku 5.5, Meta's new Muse Ember does the same work on the independent index for about a third of the money.
Muse Ember scores 36 on the Artificial Analysis Intelligence Index, the same as Haiku 5.5, at the same $0.10 input and $0.50 output list price. It costs $0.14 per index task against Haiku's $0.41 because it writes far fewer output tokens. Ember loses to Haiku on mergeable code, its price triples above 200,000 input tokens, and four of Meta's ten benchmark rows come from Meta's own runs. Re-test your routing, extraction and tool-call steps on cost per task, and keep prompts under 200K.
What Meta Released on November 9
Muse Ember is Meta's smallest multimodal model, sold as muse-ember-1.0 and built for what Meta's model page calls "high-volume agentic workflow execution". It went on sale on Monday, November 9, through the Meta Model API, OpenRouter, Google Cloud Marketplace and Meta Muse Code, Meta's terminal coding agent.
The specifications on Meta's page:
- A 1M-token context window.
- Text, image, audio and video input. Artificial Analysis lists only text and image input, so treat audio and video as unconfirmed until you test them.
- $0.10 input, $0.004 cached input and $0.50 output per million tokens for prompts up to 200K.
- $0.30, $0.012 and $1.50 per million above 200K.
The weights are closed. The Next Web reports that Ember is the first model Meta trained using its larger Muse Wonder Full, and that Meta pitches the pair as a planner and a worker: "Muse Wonder plans, Muse Ember handles the volume." Mark Zuckerberg called it "our smallest model that delivers higher intelligence per dollar than others in its class," and Meta AI chief Alexandr Wang wrote that it matches Haiku "at 3x cheaper cost per task, and smarter than luna," per the same report.
Meta's claim mostly holds up against the independent numbers, with three exceptions covered below.
Why Ember Costs Less per Task at the Same Token Price
Ember costs less because it writes fewer tokens to reach the same answer. Its token rates are no lower. All three small models list at the same rates for normal prompts: Haiku 5.5 at $0.10 and $0.50, GPT-6 Luna at $0.10 and $0.50. Cost per task is the price of the tokens a model actually burns on a piece of work, and that is where they split.
| Model (max reasoning) | AA Intelligence Index | Output tokens on the index | Cost per index task | Output speed |
|---|---|---|---|---|
| Muse Ember | 36 | 115M | $0.14 | 449 tokens/s |
| Claude Haiku 5.5 | 36 | 298M | $0.41 | 254 tokens/s |
| GPT-6 Luna | 28 | 110M | $0.15 | 102 tokens/s |
Haiku generated about 2.6 times as many output tokens as Ember to post the same score, and output is the expensive side of the bill. Luna runs about as cheaply as Ember but scores eight points lower. Artificial Analysis still calls Ember "somewhat verbose" next to the median model in its price class (115M tokens against 90M), so the saving is relative to Haiku. Ember is not unusually terse.
Latency follows the same pattern. Artificial Analysis measures Ember's time to first answer token at 4.22 seconds at max reasoning, against 180.43 seconds for Haiku 5.5 and 70.01 seconds for Luna. Those figures come from max-effort runs, and you probably don't run your routing step at max effort. They still matter for a user-facing step where someone waits on the reply.
One caution on the index itself. Fello AI's comparison notes that Meta's launch chart shows Haiku at 36.2 and Ember at 35.8, and that Artificial Analysis currently ranks Haiku first and Ember second in the price class. On the headline score, the two are tied with Haiku fractionally ahead.
Where Ember Loses: Mergeable Code and Meta's Own Benchmarks
Ember trails Haiku 5.5 on the one independent test that grades whether code is good enough to ship. On FrontierCode 2.0, which Fello AI describes as Cognition's test of whether a maintainer would merge a real repository change, Meta's own table shows Haiku at 34.9 and Ember at 27.2.
The rest of Meta's ten-row table splits into two kinds of evidence:
- Artificial Analysis runs (the rows ending in "-AA"). Ember wins AutomationBench (37.8 against Haiku's 13.9) and the Coding Agent Index (46 against 37), ties Luna on GDP.pdf long-document work (22.8), and loses FrontierCode 2.0.
- Meta's own runs. Agents' Last Exam, the Personal Superintelligence Index, OSWorld 2.1 and PerceptionBench. Per Fello AI, Meta describes its competitor runs as "best-effort" and says they may not be optimally tuned. The Personal Superintelligence Index is an internal composite of eight tests, two of which also appear separately in the table.
Terminal-Bench 4.0 is a mixed row. Fello AI reports that Ember's 40.9 comes from Artificial Analysis's harness, Haiku's 39.2 is Anthropic's self-reported Claude Code result, and Luna's 16.4 comes from the public Codex leaderboard. Three different harnesses in one column don't make a comparison, so discount that row.
The partner quotes on the model page, from Cloudflare ("up to 30% faster at the tail") and You.com (2.2-second p95 latency on coding tasks), were chosen by Meta and come without a published method, so treat them as customer references. Our guide to vendor benchmark claims explains why a vendor's own run of a rival's model is the weakest evidence on any launch page.
The practical read: Ember is a credible candidate for agent steps that move data around (route this ticket, extract these fields, call this tool, fill this form). For code a human will merge, the one independent test points to Haiku.
The 200K Tier Changes the Math on Long Prompts
Ember's price triples once a prompt passes 200,000 tokens, so the cheapest model depends on how long your prompts run. All three vendors now charge more for long prompts, and the thresholds sit in different places:
| Prompt size | Muse Ember input | Claude Haiku 5.5 input | GPT-6 Luna input |
|---|---|---|---|
| Up to 100K | $0.10 | $0.10 | $0.10 |
| 100K to 200K | $0.10 | $0.50 | $0.10 |
| 200K to 272K | $0.30 | $0.50 | $0.10 |
| Above 272K | $0.30 | $0.50 | $0.20 |
Sources: Meta, Anthropic, OpenAI. Output rises too: Ember to $1.50, Haiku to $2.50 and Luna to $0.75 per million in their long tiers.
Three things follow. Between 100K and 200K, Ember's input costs a fifth of Haiku's, because Haiku's 5x surcharge starts at 100K. Between 200K and 272K, Luna is the cheapest of the three by a factor of three on input. And above 272K, Luna's long-context rates are still lower than Ember's on both input and output.
So the 1M context window is real, but a job that routinely sends 300K tokens of contract or codebase pays Ember's long-context rate on every call. If your extraction pipeline stuffs whole documents into the prompt, the long context versus RAG math usually favours retrieving less.
What to Check Before Sending Customer Data
Read the data terms on your own Meta developer account before you route production traffic, because Meta's API has carried a training-for-discount tier before. Muse Spark's Contributor tier, reported by TechCrunch in September, cut prices by about 95% in exchange for Meta using your prompts and outputs to train future models. Ember's model page lists only the standard and long-context tiers, with no Contributor price. Check that a teammate's free trial key isn't on some other arrangement.
If you plan to buy through OpenRouter, note that its Ember listing says the model is not available through EU in-region routing. A European workload with a residency commitment can't use that route today. Our coverage of Stripe's purchase of OpenRouter covers the gateway's data-policy terms in more detail.
What to Do About Muse Ember
Your bake-off should measure cost per task on your own traffic. The three token rates are identical for most prompts, so the rate card can't separate them.
This Week:
- Pull 200 to 500 real requests from one high-volume agent step that runs on Haiku 5.5 or GPT-6 Luna today: a router, a field extractor or a tool-call step. Log the prompt length of each.
- Run them through Ember on the Meta Model API at the same reasoning setting you use in production. Record output tokens, latency and pass rate against your existing grader. The bill is the output token count, so capture it per request.
- Bucket the sample by prompt size: under 100K, 100K to 200K, over 200K. The winner can differ by bucket.
This Month:
- If Ember passes your grader on the data-moving steps, route those steps to it behind your gateway and keep Haiku or a larger model on code that a human will merge, where FrontierCode 2.0 favours Haiku.
- Get Meta's retention and training terms for your account in writing before any customer data flows, and confirm no key in your organisation is on a training-for-discount tier.
- If you already run Meta's Muse Wonder for planning, test Meta's planner-worker pairing against your current split, and score it on the cost of the whole run.
The Bottom Line
The small-model tier has turned into a price-matched market. Meta, Anthropic and OpenAI all list $0.10 and $0.50 for normal prompts, so the contest has moved to token efficiency and long-context surcharges, the same shift we saw when OpenAI halved GPT-6 Sol's token price and buyers learned the reasoning setting mattered more than the rate card. On the index, Ember wrote about 61% fewer output tokens than Haiku for the same score, and that gap drives most of the price difference.
Ember's case is strongest on the independent rows and weakest on the rows Meta ran itself. Price it on your own traffic, keep prompts under 200K tokens, and keep merge-ready code on the model that wins FrontierCode.
Continue Reading
- Claude Haiku 5.5 Charges 5x for Prompts Over 100K Tokens
- Meta's Muse Wonder Trails Opus 5.5 by 6 Points at Same Price
- GPT-6 Sol Halves the Token Price but Benchmarks It at Top Effort
- Vendor AI Benchmarks Need Independent Runs With Task Logs
- Long Context vs RAG Cost: Cached, Each Answer Still Costs 5x
- Meta Releases Muse Spark Open With Critical Infrastructure Ban