Muse Ember has the same $0.10/$0.50 card as Haiku 5.5, and the same score. An AA task costs $0.14 on Ember and $0.41 on Haiku, and output tokens explain only a quarter of the gap.
Three small models now charge $0.10 in and $0.50 out. Meta's is the one that comes out cheapest per task, and the reason is a line on the rate card most people skip.
Image source: Meta
Meta released Muse Ember on November 9. It's the smallest model in the Muse family and the first one Meta trained using Muse Wonder Full. The API ID is muse-ember-1.0, the window is 1M tokens, and it can write up to 512K tokens in one response.
The price is $0.10 per million input tokens and $0.50 per million output. Claude Haiku 5.5 and GPT-6 Luna charge exactly that. So on the pricing table the three look identical. On Artificial Analysis they don't: Ember scores 36.0 on the Intelligence Index for $0.136 a task, and Haiku 5.5 scores 36.5 for $0.411.
Fewer output tokens explain some of that. Not most of it. Most of the gap is cached input, which is where small models doing agent work spend their money, and where Meta undercut everyone.
Same headline price, different fine print
| Per 1M tokens | Muse Ember | Haiku 5.5 | GPT-6 Luna |
|---|---|---|---|
| Input | $0.10 | $0.10 | $0.10 |
| Cached input | $0.004 | $0.01 | $0.01 |
| Cache write | none listed | $0.125 (5 min) | $0.125 (30 min) |
| Output | $0.50 | $0.50 | $0.50 |
| Long tier starts at | 200K | 100K | 272K |
| Long tier in / cached / out | $0.30 / $0.012 / $1.50 | $0.50 / $0.05 / $2.50 | $0.20 / $0.02 / $0.75 |
| Batch in / out | none | $0.05 / $0.25 | $0.05 / $0.25 |
| Max output | 524,288 | 128K | 128K |
From Meta's pricing page, Anthropic's pricing page and the GPT-6 Luna model page, November 11, 2026. Every long tier applies to the whole request, not just the tokens past the line.
Two rows decide most agent bills. A cached token on Ember costs 4% of a fresh one. On Haiku and Luna it's 10%. And Haiku's long tier starts at 100K, where the whole request jumps to 5x on input and output. An agent that has been reading files for twenty turns crosses 100K quickly. Ember gives you twice the room before its price goes up, and when it does go up, the step is 3x.
Meta counts cached tokens toward the 200K line, including the stored conversation when you continue with previous_response_id. Its own system prompt and steering text are added to every request and not billed.
Where the $0.27 goes
AA's Ember page splits the $0.136 into $0.094 of input of every kind and $0.042 of output. Inside the input, $0.066 is cache reads, $0.027 cache writes and well under a cent fresh input. Output is $0.026 of reasoning and $0.016 of answer. So about 69% of an Ember task is reading, not writing.
Now Haiku. It wrote 297.7M output tokens across the index to Ember's 114.9M, about 219K per task against 84.5K. Even if all of Haiku's requests stayed under 100K, that output would cost about $0.11 a task, against Ember's $0.04. That accounts for $0.07 of a $0.27 gap.
The remaining $0.21 is input, nearly all of it cached context, billed at 2.5x Ember's rate below 100K and 12.5x Ember's short rate above it. AA doesn't publish Haiku's per-tier split, so we can't say exactly how much came from the long tier. The direction is clear enough.
One thing we'd flag on Ember's side too. AA counts about 7.4M input tokens per task. At $0.004 that would be $0.03, yet AA bills $0.094 for input. We read that as a good share of requests crossing 200K and paying $0.012 cached and $0.30 fresh, plus a cache write charge Meta doesn't list. Plan with $0.14, not with the card.
Cost per task, small-model tier
Cost per Artificial Analysis Intelligence Index task in USD, v5.0.1, read November 11, 2026. Muse Wonder max (43.6, $6.95) is left off so the small bars stay readable.
Ember on xhigh beats GPT-6 Luna on max by 6 points and costs less. Ember on max matches Haiku on max for a third of the money. Haiku on xhigh lands half a point above Ember xhigh at twice the price. Meta sets max as the default, and unusually, the default is the setting we'd pick.
Run 10,000 tasks like these a month and Ember costs about $1,360. Haiku 5.5 on max is about $4,110. Batch would roughly halve Haiku's bill, and Ember has no batch tier. Even then Haiku would be around $2,050, half again Ember's.
It's also faster: 454 output tokens a second on AA against Haiku's 255, and 3.1 minutes per task.
It passes on 64% of trivia questions
On AA-Omniscience, Ember gets 21.1% of questions right. Haiku gets 34.7%. AA says Ember declines 64% of questions, and its hallucination rate is 17.4% against Haiku's 34.2%. For an agent that's often what you want. For a chatbot answering trivia from memory, it isn't.
The fix Meta sells is web search at $2.50 per 1,000 queries. Price that against the tokens. A 10K-in, 1K-out request costs $0.0015 on Ember. One search adds $0.0025. On a model this cheap, a single search more than doubles the request, so cap the tool calls.
Coding is a split. Ember wins Terminal-Bench 4.0 on AA, 40.9% to Haiku's 32.8%, and loses FrontierCode 2.0, 27.2% to 34.9%. Meta's own chart shows Haiku at 39.2 on Terminal-Bench, because Meta ran each model in its best harness, Claude Code for Haiku. Use AA's numbers when comparing.
68.1 or 37.8?
AA's launch article gives Ember 68.1 on AutomationBench, and Meta's post gives 37.8. Both are right. AA's AutomationBench page reports a partial-credit score as the headline and a strict score beside it. Meta quoted the strict one.
| Model | Partial credit | Strict |
|---|---|---|
| Muse Ember max | 68.1 | 37.8 |
| Muse Wonder max | 68.1 | 39.1 |
| GPT-6 Astra max | 68.5 | 41.6 |
| GPT-6 Luna max | 53.2 | 20.2 |
| Claude Haiku 5.5 max | 35.4 | 13.9 |
Artificial Analysis AutomationBench, read November 11, 2026.
On partial credit Ember ties Wonder, which costs 40x more per token. On strict scoring it trails Wonder by 1.4 points. Either way it's far ahead of the two models with the same price card, and that's the comparison that matters for the budget.
Inside Meta's own lineup
Muse Spark 1.3 is $1.25 in, $4.25 out and $0.15 cached. On AA it scores 36.6 at max for $2.39 a task, half a point over Ember for 17.5x the money. Unless you need something Spark does that the index doesn't measure, Ember replaces it.
Spark's Contributor tier is $0.10 in and $0.20 out, cheaper than Ember on output, but Meta trains on what you send and caps it at 100 requests a minute. Ember has no Contributor tier. Standard accounts get 3,000 requests and 4M tokens a minute per team.
Muse Wonder is the step up at 43.6 for $6.95 a task. Our Wonder post covers it. Ember is not open weights; Meta's open models are still Muse Spark Open and Muse Glimmer.
Should you move off Haiku?
If you run a small-model agent on Haiku 5.5 or Luna today, Ember is worth an eval this week. Same headline price, about a third of Haiku's per-task cost on AA, and twice Haiku's long-context headroom before the price steps up. We'd still keep Haiku for code generation, where FrontierCode favours it, and for anything that has to go through batch.
Watch your prompt sizes either way. On all three models, a request that tips over the long-context line pays the higher rate on every token in it, not just the extra ones. Log input tokens per request and see how often you cross 100K or 200K before trusting any per-token comparison, including ours.
Ember has its own model page. Line it up against Haiku 5.5 and Luna on the pricing table, or put your own cache hit rate into the calculator. Our Haiku 5.5 post explains its 100K tier in more detail.
Sources
- Meta: Introducing Muse Ember - release, lineage, Meta-run benchmarks and harness notes
- Meta pricing and rate limits and Muse Ember model doc - rates, 200K tier rules, web search, rate limits, ID, context, max output, effort default
- Artificial Analysis launch article and model page - index scores, cost per task and its split, tokens, speed, Omniscience; comparison rows from each model's page on the same site
- Artificial Analysis AutomationBench - partial and strict scores
- Anthropic pricing and GPT-6 Luna model page - comparison rate cards and long-context tiers