Muse Ember: Pricing, Benchmarks, and Haiku 5.5 Verdict

Four weeks after Muse Wonder, Meta shipped the other end of its lineup: its smallest model. Muse Ember launched on November 9, 2026 at $0.10 per million input tokens and $0.50 per million output tokens, the same list price as Claude Haiku 5.5 and GPT-6 Luna on standard-length prompts.

Meta's headline claim is "more intelligence per dollar" than either rival at any reasoning setting. That mostly held up when Artificial Analysis ran its own tests, with one catch worth knowing before you switch: Ember ties Haiku 5.5 on overall intelligence, wins most of the agent and tool-use tests, and loses on hard coding and factual recall.

I went through Meta's launch post, Artificial Analysis's write-up and model page, and our own Muse Wonder breakdown to figure out what Ember is actually good at, what it costs once you count tokens, and where the cheaper option is the wrong call.

TL;DR

• What it is: Meta's smallest, most efficient Muse model, built for high-volume work like tool calls, coding subtasks, routing, data extraction, and long-document parsing.
• Price: $0.10 input, $0.004 cached input, and $0.50 output per million tokens for prompts up to 200K input tokens; $0.30, $0.012 and $1.50 above that. A 1M-token context window and up to 512K output tokens.
• Intelligence: Scores 36 on the Artificial Analysis Intelligence Index at max reasoning, matching Claude Haiku 5.5 and sitting one point below Muse Spark 1.3.
• Cost per task: About $0.14 per Artificial Analysis task, against roughly $0.41 for Haiku 5.5 at the same score. That's about a third.
• Wins: Agentic workflows, terminal tasks, tool-assisted perception (Meta's own eval), and coding inside Meta's own Muse Code agent.
• Losses: FrontierCode 2.0 (27.2% versus Haiku's 34.9%), knowledge work, and factual recall. It declines to answer far more often than its rivals.
• Where to get it: Muse Code and the Meta Model API as muse-ember-1.0. Meta says it's also on Google Cloud and Oracle Cloud.

What is Muse Ember?

Muse Ember is Meta's smallest and most efficient Muse model, released November 9, 2026, and built for the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long-document parsing.

Meta splits the family like this: Muse Wonder for the hardest problems, Muse Spark for coding and agentic work, and Muse Ember for repeatable tasks at about $0.14 per Artificial Analysis task. Meta's launch post puts the pairing in one line: "Muse Wonder plans, Muse Ember handles the volume."

According to Meta, Ember is natively multimodal across text, image, audio, and video, and it's the first model Meta trained using Muse Wonder Full.

Meta also says Ember was evaluated under its Superintelligence Scaling Framework for chemical and biological, cybersecurity, and loss-of-control risks, and came in below the high-risk thresholds there. Separately, Meta says it tested how Ember behaves when acting on a user's behalf, including resisting instructions hidden in pages or tool outputs and checking with a human before high-impact actions. It's a small model, but Meta says it got the same core safety training as the big ones.

Where Ember sits in the Muse lineup

Ember is the cheap executor at the bottom of the stack. Here's the lineup on price and on the Artificial Analysis Intelligence Index.

The lineup at a glance:

Wonder's pricing and Wonder Full's access come from our Muse Wonder article. The index numbers are Artificial Analysis's current scores at max reasoning. Our earlier Muse Wonder piece showed an older score of 43 for Wonder.

The Spark line is the one to stare at. At max reasoning, Ember scores 36 to Spark 1.3's 37, but costs $0.14 per task against $2.39. That's a one-point drop for a bill about 17 times smaller. For most repeatable work, one point is noise and the price is the whole story.

Muse Ember pricing

All three efficient-tier models list at the same headline rate on standard prompts: $0.10 per million input tokens and $0.50 per million output tokens. Two of them charge more for long prompts. Ember triples to $0.30 input and $1.50 output above 200K input tokens, and Haiku 5.5 jumps to $0.50 and $2.50 above 100K. On the Meta Model API, Ember adds $0.004 per million cached input tokens, a 1M-token context window, and up to 512K output tokens.

List price, side by side:

Ember's prices come from Meta's launch and developer posts. Haiku 5.5 and Luna list prices are as reported by Artificial Analysis; their cached-input and long-prompt rates come from the vendors' own pricing pages. For the rest of Anthropic's lineup, see our Claude pricing breakdown.

Matching input and output prices still leave very different per-task bills. The gap shows up in how many tokens each model burns to finish a job.

Cost per task, side by side:

Those are Artificial Analysis figures. Meta's own chart uses slightly different values ($0.136 and 35.8 for Ember at max, $0.411 and 36.2 for Haiku at max), which is why the 36-to-36 tie in the table below is a rounded one, and its captions are a little slippery: one version labels the Ember costs "estimates at proposed pricing" and another says "list pricing," so take Meta's cost chart as an estimate, not a quote.

Two things make Ember cheaper per task:

• It uses fewer output tokens. Artificial Analysis found Ember gets to the same score as Haiku 5.5 at max using roughly a third as many output tokens.
• Its cached input costs 60% less. At $0.004 per million tokens against $0.01 for Haiku 5.5 and Luna, Ember is cheaper for an agent that resends the same system prompt and tool definitions on every call, and Artificial Analysis says cached input is most of the cost in its agent tests.

Artificial Analysis also put Ember at about 4 times cheaper per task than DeepSeek V4.1 Flash (33 on the index, $0.60 per task) and GLM-5.3-Flash (28, $0.58).

Muse Ember benchmarks

Ember wins the agent-style tests and drops a few important ones, so I've split the numbers into three buckets: Meta's own table, what Artificial Analysis found running its own tests, and what launch partners said.

Meta's table: Ember vs Haiku 5.5 vs GPT-6 Luna

All three at max reasoning, as reported by Meta. Several rows are Meta's internal evaluations, so I've labeled them.

The pattern is clear enough. Ember's wins cluster on agent-style work: multi-step business workflows, terminal tasks, and operating a computer. On the OSWorld partial score and Terminal-Bench, its lead over Haiku is under two points. On pure coding quality, measured by FrontierCode, Haiku 5.5 wins by 7.7 points.

What Artificial Analysis found

Artificial Analysis ran its own tests and mostly backed up Meta's picture, with a few differences worth knowing:

• Overall: Ember is the cheapest model scoring 36 on the Intelligence Index and sits on AA's cost-versus-intelligence frontier. It trails Muse Wonder by 8 points at max reasoning.
• AutomationBench-AA: AA reports 68.1% for Ember, level with Muse Wonder and GPT-6 Astra. Meta's chart shows 37.8%. The two aren't the same measurement. Artificial Analysis publishes AutomationBench-AA under two metrics: "Score," the average share of each task's objectives completed, zeroed out by any guardrail violation, and "Tasks Completed," the share of tasks finished in full. AA's write-up headlines the Score; Meta's chart shows the stricter Tasks Completed figure, counting only first-try passes. Read each number against its own metric, not against the other.
• Agentic coding: inside Muse Code, Ember scores 46 on AA's Coding Agent Index, against 37 for Haiku 5.5 in Claude Code and 31 for Luna in Codex. Each model ran in its own vendor's agent, so this partly measures the harness too.
• Terminal-Bench 4.0: 40.9% for Ember versus 32.8% for Haiku in AA's runs. Meta's table lists Haiku at 39.2% using the best published harness, so the size of Ember's lead depends on the setup.
• Knowledge work: Ember trails Haiku on AA-Briefcase, 1521 Elo to 1574.
• Factual recall: On AA-Omniscience, Ember answered 21.1% of questions correctly, against 34.7% for Haiku 5.5 and 50.5% for Gemini 3.8 Flash. It declined to answer 64% of questions, which kept its hallucination rate at 17.4%, about half of Haiku's.

That last point cuts both ways. Ember is less likely to make things up, but it also knows less and says "I don't know" a lot. For extraction and tool calls, where the answer is sitting in the input, that's fine. For questions that lean on what the model remembers, it's a real weak spot.

What early partners reported

Meta published quotes from several launch partners. These are vendor-selected testimonials, not independent tests, so weigh them accordingly:

• Cloudflare: up to 30% faster at the tail than the comparable model it tested.
• You.com: a p95 latency of 2.2 seconds, 4 to 13 times faster than leading open-weight models, and 36% to 48% fewer output tokens.
• Box, Cline, and Oracle gave quotes too, on contract scoring, subagent coding, and engineering tasks.

‍

{{templates}}

‍

How to access Muse Ember

Ember is available today through Meta's API and Meta's coding agent, and Meta says it's also on Google Cloud and Oracle Cloud.

Where Ember lives:

On the API, Ember costs $0.10 per million input tokens, $0.004 cached, and $0.50 output for prompts up to 200K input tokens, and $0.30, $0.012 and $1.50 above that, with a 1M-token context window and up to 512K output tokens. Ember is included in all four Muse Code plans, from the $5 Everyday tier to the $200 Maximum tier, so you don't need a plan change to try it.

Where Muse Ember falls short

The cheap model is not the better model everywhere. Four places to watch:

1. Harder coding. Haiku 5.5 beats Ember on FrontierCode 2.0 by 7.7 points. If you're wiring up coding agents that need to write merge-ready code on their own, Haiku is ahead.
2. Factual recall. Ember gets about 21% of AA-Omniscience questions right, well behind Haiku 5.5 and Gemini 3.8 Flash. Good for extraction, risky for anything that leans on general knowledge.
3. Knowledge work. Haiku 5.5 leads on AA-Briefcase, Artificial Analysis's test of long business tasks that end in spreadsheets, slides, and memos.
4. Some headline wins are Meta's own. Agents' Last Exam, the Personal Superintelligence Index, and the OSWorld numbers in Meta's table are internal evaluations, so treat them as Meta's scoring until someone independent reruns them.

For hard single-shot coding and knowledge-heavy questions, Meta's own docs point you to Muse Spark or Muse Wonder, and among small models, Haiku 5.5 does better on both.

Muse Ember vs the alternatives

If you're choosing among the cheap tier, the right pick depends on the job. GPT-6 Luna doesn't get a row: at the same list price it costs about the same per task as Ember ($0.15 against $0.14) and scores 8 points lower on the index, so nothing in these results points to it.

Which model for which job:

For more small-model options, see our breakdowns of Gemini 3.8 Flash and Qwen 3.8.

Should you use Muse Ember?

Use Muse Ember if most of your work is repeatable: classifying tickets, pulling fields out of documents, routing requests, calling tools in a fixed workflow, or running as a subagent under a bigger model. That's the job Meta built it for, and the cost-per-task numbers back it up.

Skip it as your only model if you need strong standalone coding, broad general knowledge, or polished knowledge-work output. Haiku 5.5 lists the same input and output rates on standard prompts and does better on all three, though it costs about three times as much per task.

The bigger takeaway is the shape of the release. Labs now ship a planner and a cheap executor as a pair, and the savings come from sending each task to the right one. If your team works in Slack, Lindy is an AI teammate that's model agnostic, so you can match the model to the job instead of pushing everything through one.

Match the model to the job

The planner-and-executor split is nice in a benchmark table and awkward in real life. Someone still has to decide which model handles which task, wire it up, and redo that work every time a cheaper model lands (which, lately, is most weeks).

That's the part Lindy takes off your plate. Lindy is an AI teammate that works in Slack and executes the tasks you hand it, and it can ask for your OK before anything with outside impact goes out.

With Lindy, teams can:

• Pick the model for each step from most major models: a budget model for the high-volume steps, a frontier model for the hard ones.
• Text it to handle inbox triage, scheduling, and follow-ups, with an approval step before anything goes out if you want one.
• Work where the team already is, across Slack, email, and text.
• Switch models when a new one pulls ahead, so the next model launch is a settings change.

Say you tell Lindy in Slack to triage the support inbox on a budget model and hand anything that needs a real answer to a frontier model. It sorts the queue, drafts the replies, and checks in with you before anything goes out.

Match the model to the job and a launch-day release is a dropdown change.

Try Lindy free and let an AI teammate handle the repeatable work, then pick the model that fits once you're on a Team plan.

‍

{{cta}}

‍

Frequently asked questions

What is Muse Ember?

Muse Ember is Meta's smallest and most efficient Muse model, released November 9, 2026. Meta built it for high-volume work like tool calls, coding subtasks, routing, extraction, and long-document parsing. It's available through the Meta Model API as muse-ember-1.0 and in Muse Code, and Meta says it's also on Google Cloud and Oracle Cloud.

How much does Muse Ember cost?

Muse Ember costs $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens for prompts up to 200K input tokens, per Meta; above that it is $0.30, $0.012 and $1.50. Artificial Analysis measured about $0.14 per task on its Intelligence Index at max reasoning, roughly a third of Haiku 5.5's cost at the same score.

Is Muse Ember better than Claude Haiku 5.5?

No, not across the board. Muse Ember ties Claude Haiku 5.5 at 36 on the Artificial Analysis Intelligence Index for about a third of the cost per task, and it wins most of the agentic benchmarks in Meta's table. Haiku 5.5 is still ahead on FrontierCode 2.0 coding, AA-Briefcase knowledge work, and factual recall.

Is Muse Ember better than GPT-6 Luna?

Yes, on most published results. Muse Ember beats GPT-6 Luna on the Artificial Analysis Intelligence Index (36 to 28) and on every benchmark in Meta's table except GDP.pdf, where they tie at 22.8%. Both list at $0.10 input and $0.50 output per million tokens.

What's the difference between Muse Ember and Muse Wonder?

Muse Wonder is Meta's model for the hardest problems, priced at $4 per million input tokens and $20 per million output tokens. Muse Ember costs $0.10 and $0.50 and scores 8 points lower on the Artificial Analysis Intelligence Index. Meta pitches them as a pair: Wonder plans, Ember handles the volume.