Meta Muse Ember: Same Price as Claude Haiku, a Third of the Cost per Task

Muse Ember is Meta’s smallest and cheapest AI model, released on November 9, 2026. It is built for the small, repeated steps AI agents run all day: tool calls, routing, data extraction and quick code fixes. It lists at $0.10 per million input tokens and $0.50 per million output tokens, exactly what Claude Haiku 5.5 and GPT-6 Luna cost. The difference is how many tokens it burns. Artificial Analysis scores Meta Muse Ember 36, level with Claude Haiku 5.5, at $0.14 per task against $0.41.

That makes the headline claim true but easy to misread. Muse Ember is not cheaper per token than its rivals. Above 200,000 tokens of context its price triples, a detail that sits on Meta’s developer page but not in the launch post. Below: the full benchmark table with who actually ran each test, the two price tiers, where Ember fits next to Muse Spark and Muse Wonder, and whether you can use it.

The Key Takeaways

• Same sticker price: Muse Ember, Claude Haiku 5.5 and GPT-6 Luna all list at $0.10 in / $0.50 out per million tokens for normal-length prompts.
• Cheaper per task: on the Artificial Analysis Intelligence Index, Ember scores 36 for $0.14 a task; Haiku 5.5 scores 36 for $0.41, because it wrote 298M output tokens to Ember’s 115M.
• Wins 7 of Meta’s 10 rows, ties 2 and loses 1: Claude Haiku 5.5 beats it on FrontierCode 2.0, the test of whether code is good enough to merge (34.9 vs 27.2).
• The 200K catch: prompts over 200,000 tokens cost $0.30 in / $1.50 out, three times the headline rate.
• Developers only: Muse Ember runs in the Meta Model API and Muse Code. Meta has not put it in the Meta AI app or the Muse agent.

What Is Meta Muse Ember?

Meta Muse Ember is the smallest model in Meta’s Muse family, built to be the cheap worker inside an AI agent rather than the model you chat with. Meta Superintelligence Labs released it on November 9, 2026, with the API model ID muse-ember-1.0. Meta’s launch post describes it as designed for “the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing.”

Muse Ember has a 1 million token context window and can write up to 512,000 tokens in one answer. Meta says it natively understands text, images, audio and video. An agent can read a screenshot, a scanned contract or a meeting recording without handing it to a second model. One discrepancy is worth knowing: the Artificial Analysis model page lists Ember’s inputs as text and image only, so audio and video support rests on Meta’s word for now.

Muse Ember is also the first Meta model trained using Muse Wonder Full, the larger frontier model Meta released in October. In practice that means the big model taught the small one. Meta’s own pitch for how to use the two together fits in one line: “Muse Wonder plans, Muse Ember handles the volume.”

AI at Meta announced Muse Ember on X with the same two claims the rest of this article tests, efficiency and agent work:

Muse Ember Benchmarks: The Full Table and Who Ran Each Test

Muse Ember leads or ties on 9 of the 10 benchmarks Meta published against Claude Haiku 5.5 and GPT-6 Luna, all three at max reasoning. The table below adds the column Meta’s chart leaves out: who produced each number. Five rows come from Artificial Analysis, an independent testing lab, four are Meta’s own runs, and one mixes sources.

Higher is better in every row. Scores are percentages except the two indexes. Sources: Meta’s launch post and evaluation report, Artificial Analysis, November 9, 2026.

Where Muse Ember wins

Muse Ember’s clearest win is business automation. On AutomationBench-AA, Artificial Analysis’s run of Zapier’s 657 workflow tasks across simulated apps like Gmail, Salesforce and Slack, Ember completes 37.8% of tasks in full. Claude Haiku 5.5 manages 13.9%, so Ember finishes 2.7 times as many. That is the exact job Meta built it for: pull a record, call a tool, update another app, repeat.

Muse Ember also leads the Coding Agent Index-AA, 46 to 37 over Haiku 5.5. That comparison is fairer than most, because Artificial Analysis ran each model inside its maker’s own coding agent: Muse Code for Ember, Claude Code for Haiku and Codex for Luna. Against GPT-6 Luna, Ember wins every row except GDP.pdf, where they tie. The biggest gap is Terminal-Bench 4.0, at 40.9 against 16.4.

Where Claude Haiku still wins

Claude Haiku 5.5 beats Muse Ember on FrontierCode 2.0, 34.9 to 27.2. FrontierCode, from Cognition, grades real changes in real repositories on whether a maintainer would merge them. That makes it the closest row to asking whether the code is good enough to ship. If your agent writes pull requests rather than small fixes, that row matters more than the others.

Haiku 5.5 also edges Muse Ember on the unrounded Artificial Analysis Intelligence Index. Meta’s own cost chart, built on the index as published on launch day, shows 36.2 against 35.8. Artificial Analysis’s live data, now on index version 5.0.1, puts them at about 36.5 and 36.0. Both round to 36, and Artificial Analysis ranks Claude Haiku 5.5 first and Ember second among the 32 models in their price class. “Matches Haiku” is fair. “Beats Haiku” on the index is not.

How Meta picked the numbers

Meta’s four-page evaluation report explains two choices that flatter Muse Ember. First, the Terminal-Bench 4.0 row mixes sources. Ember’s score is an Artificial Analysis run in the mini-swe-agent harness. Haiku’s is Anthropic’s self-reported Claude Code result, and Luna’s is its Codex result from the public leaderboard. Three different setups sit in one row.

Second, Meta ran four of the ten rows itself. Its report calls the competitor runs “best-effort” and warns that prompts, tools and runtime “may not be optimally tuned for proprietary models.” The Personal Superintelligence Index is an internal composite of eight tests, and two of them, AutomationBench and GDP.pdf, already appear in the table on their own. Read the five Artificial Analysis rows as the independent evidence, and the rest as Meta’s case.

Muse Ember Pricing: Same Sticker as Haiku and Luna, Fewer Tokens

Muse Ember costs $0.10 per million input tokens and $0.50 per million output tokens on the Meta Model API. That is the same list price as Claude Haiku 5.5 and GPT-6 Luna for normal-length prompts. TechCrunch’s launch story called Ember “significantly cheaper than frontier competitors like Haiku and Luna.”

Per token, it is not. The one list-price difference is cached input: $0.004 per million tokens for Ember against $0.01 for both rivals, which matters for agents that resend the same long instructions on every call.

Prices from Meta’s developer page, Anthropic’s pricing docs and OpenAI’s API pricing page, checked November 9, 2026. Cost per task is Artificial Analysis’s average cost to run one Intelligence Index task at max reasoning.

Why Muse Ember costs less per task

Muse Ember costs less per task because it writes fewer tokens to reach the same answer. Running the full Artificial Analysis Intelligence Index, Ember at max reasoning produced 115 million output tokens. Claude Haiku 5.5 at max produced 298 million, about 2.6 times as many, at the same $0.50 output rate. That gap is most of the difference between $0.14 and $0.41 a task.

Muse Ember is not unusually terse, though. Artificial Analysis calls it “somewhat verbose” next to the median model in its price class, which wrote 97.1 million tokens. Ember simply talks far less than Haiku 5.5 at max. GPT-6 Luna wrote 110 million tokens and costs about the same per task, $0.15, but scores 8 points lower. If you want to understand why token counts drive the bill, our guide to GPT-6 Sol and GPT-6 Luna pricing walks through the same arithmetic for OpenAI’s models.

The 200K price step

Muse Ember’s price triples once a prompt passes 200,000 tokens: $0.30 per million input, $0.012 cached and $1.50 output. Meta’s launch post quotes only the lower rate, while the pricing table on Meta’s Muse Code developer page lists both. Feeding Ember a whole codebase or document set, the use Meta advertises for the 1M window, lands on the higher tier.

Even so, Muse Ember stays cheaper than Claude Haiku 5.5 on long prompts. Haiku 5.5 switches to $0.50 in and $2.50 out above 100,000 tokens. Between 100K and 200K, Ember costs a fifth of Haiku per input token, and above 200K it still costs less.

The fair summary: same price on short prompts, cheaper on long ones, and cheaper per task across the board.

Muse Ember vs Muse Spark vs Muse Wonder

Muse Ember is the bottom tier of Meta’s three-model lineup. Muse Wonder takes the hardest problems, Muse Spark takes coding and agent workflows, and Muse Ember takes repeated agent steps at scale. The surprise is how close Ember sits to Muse Spark 1.3 on the Artificial Analysis index for a fraction of the cost.

Muse Ember scores one point below Muse Spark 1.3 and costs about 17 times less per task, $0.14 against $2.39. For most agent subtasks that trade is easy. Meta’s own pricing table hides one wrinkle: the Contributor version of Muse Spark 1.3 lists at $0.10 in and $0.20 out, cheaper per output token than Ember. The catch is in its label, “Used to improve our products,” so anything you send it can train Meta’s models. Ember carries no such label.

For the planner side of the pairing, Meta Muse Wonder costs $4 in and $20 out per million tokens, 40 times Ember’s input price. That price gap is the whole argument for Meta’s planner-and-worker setup: let the expensive model decide once, and let Ember carry out the hundred small steps that follow.

How to Use Muse Ember Today

You can use Muse Ember in two places today: the Meta Model API, where you pay per token, and Muse Code, Meta’s terminal coding agent. Meta’s developer playground also lists the model for signed-in developers. Meta has not put Ember in the Meta AI app, WhatsApp or its consumer Muse agent. The launch post only says Ember’s cost is “low enough to run in every product, for everyone.”

Muse Ember on the API and in Muse Code

Developers have two routes to Muse Ember, and both need an account with Meta’s developer platform:

1. API: sign in at Meta’s developer site, create a key and call muse-ember-1.0. Meta says the API works with OpenAI-SDK-compatible clients, and Ember supports the Responses API.
2. Muse Code: install the command-line agent and pick a plan: Everyday Usage at $5 a month, High Usage at $15, Power Usage at $50 or Maximum Usage at $200. The plan cards on November 9 still listed Muse Spark, Muse Wonder and Muse Image, not Ember, so check the model picker after you sign in.

Muse Ember for everyone else

For most people, Muse Ember is not something to download. It is a part other companies will build into their agents. If what you wanted was a fast, capable model to talk to, you don’t have to bet on one vendor. Fello AI puts Claude, GPT, Gemini, Grok and DeepSeek in one Mac, iPhone and iPad app, so you switch models in the same chat instead of opening a new account.

Mark Zuckerberg posted the launch about ten minutes before Meta’s AI account did, and his version names the same two places to run it:

Should You Switch Your Agent’s Subtasks to Muse Ember?

Yes, test Muse Ember if your agent spends most of its tokens on routine steps: tool calls, extraction, routing, document reading and small code fixes. On Artificial Analysis’s independent numbers it matches Claude Haiku 5.5’s score at a third of the cost per task, and it beats GPT-6 Luna at about the same cost. Partner testing points the same way. Every partner quote in Meta’s post was chosen by Meta, but the numbers are specific: Cloudflare’s CTO reports Ember ran “up to 30% faster at the tail,” and You.com measured a 95th-percentile latency of 2.2 seconds on coding tasks.

Keep Claude Haiku 5.5 or a bigger model where the output has to be merge-ready code. FrontierCode 2.0 is the one independent test Ember loses, and it is the one closest to a real pull request. Watch your prompt sizes, too: past 200,000 tokens, Ember’s rate triples.

The verdict: Muse Ember is not cheaper than its rivals on the price list, and Meta never quite says it is. It is cheaper on the bill, because it reaches Haiku-level answers with about 60% fewer output tokens. For agents that run thousands of small steps a day, the bill is the number that matters. On that number, Muse Ember beats both rivals Meta chose to compare it with.

Frequently Asked Questions