Meta’s proprietary Muse Ember nearly matches Claude Haiku 5.5 at 1/3 the estimated per-task cost

Meta released Muse Ember on Monday, giving developers a cheaper model for the work that can make AI agents expensive to run at scale.

A single support ticket, for instance, can send an agent through a series of lookups, document reads and tool calls before it produces an answer. Ember is built to handle those steps, as well as jobs like running tests and making small code fixes.

In independent evaluation firm Artificial Analysis’s launch evaluation, Ember scored 36 on the firm’s Intelligence Index at maximum reasoning effort, the same rounded score as Anthropic’s Claude Haiku 5.5. Ember cost about 14 cents per benchmark task, compared with 41 cents for Haiku.

Meta’s underlying chart puts the unrounded scores at 35.8 and 36.2, respectively. OpenAI’s GPT-6 Luna scored 27.8 at about 15 cents per task at its highest setting, although it ran for as little as about 7 cents at a lower setting with a score of 21.4.

Meta CEO Mark Zuckerberg described Ember in a post on X as “our smallest model that delivers higher intelligence per dollar than others in its class.”

“[Muse W]onder can plan, ember can handle the volume,” Meta Chief AI Officer Alexandr Wang wrote in his launch thread on X.

A million-token window with a 200,000-token price jump

The Meta Model API offers the proprietary model as muse-ember-1.0 through its Responses, Chat Completions and Messages endpoints.

It accepts text, images, audio, video and PDFs and returns text. Meta lists a context window of 1,048,576 tokens and a maximum output of 524,288 tokens. Reasoning effort ranges from minimal to max, with max as the default. Ember is also available in Muse Code. Meta lists Oracle Cloud as another route, while Google Cloud’s preview documents text and image inputs.

At up to 200,000 input tokens, Meta charges $0.10 per million uncached input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens for Muse Ember.

Above 200,000, the rates rise to $0.30, $0.012 and $1.50, respectively, matching.

The listed API pricing makes Muse Ember among the cheaper foundation-class or near-foundation models available globally, matching the uncached sub-200K token price of Anthropic's Claude Haiku 5.5 and OpenAI's GPT-6 Luna.

The higher rates apply to the entire request, including its output. Cached input and conversation history carried forward through previous_response_id count toward the cutoff. For a request with 200,000 uncached input tokens and 10,000 billed output tokens, the model charge is 2.5 cents. One more input token takes that same request to about 7.5 cents.

Anthropic’s standard API pricing starts at the same $0.10 input and $0.50 output for Haiku, then rises to $0.50 and $2.50 above 100,000 input tokens.

Its cache reads start at $0.01 per million. OpenAI’s standard pricing lists the same starting input and output rates for Luna, with $0.01 for cached input. Its long-context rates are $0.20 input, $0.02 cached input and $0.75 output.

A workload with large prompts and long answers can therefore favor OpenAI's GPT-6 Luna on token rates, even when Ember comes out cheaper on the benchmark suite.

Artificial Analysis attributes part of Ember’s benchmark cost advantage to cheaper cache reads and reports that Ember used roughly a third as many output tokens as Haiku at maximum effort.

That distinction matters when estimating a production bill. The same list price can yield very different costs once an agent starts reasoning, retrieving records and carrying the results from one step into the next.

Ember’s clearest lead is in business workflows

In Meta’s launch results, Ember fully completed 37.8% of AutomationBench-AA tasks on its first attempt, compared with 20.2% for Luna and 13.9% for Haiku. The test puts agents to work across simulated business apps.

Artificial Analysis’s headline measure, which awards credit for individual objectives completed without violating the benchmark’s rules, puts Ember at 68.1%. The two numbers answer different questions. For a workflow that only counts when every step is complete, the first figure is the relevant one.

The coding results are more mixed. In Artificial Analysis’s Coding Agent Index, Ember in Muse Code scores 46, ahead of Haiku in Claude Code at 37 and Luna in Codex at 31.

But Haiku leads on FrontierCode 2.0, which grades whether changes are ready to merge, at 34.9% to Ember’s 27.2%. And in Meta’s desktop-app evaluation, OSWorld 2.1, Ember’s 69.5% partial-credit score falls to 33.8% when tasks are graded on an all-or-nothing basis.

Together, those results make a stronger case for giving Ember a bounded job with a clear check than for handing it an entire engineering ticket or a desktop workflow unattended.

How the open-weight alternatives stack up

Ember’s closest comparisons are by workload and price. Meta has kept Ember’s weights and parameter count private, so its size cannot be matched directly to models with downloadable weights. The scores below use Artificial Analysis’s current v5 methodology. The firm revised the index in October, so some scores in earlier launch coverage are higher and belong to the previous version.

DeepSeek-V4.1-Flash-1017 is a close alternative for inexpensive hosted agent work with the option to run the MIT-licensed weights yourself. DeepSeek’s API offers a million-token window and image input, and charges $0.15 input and $0.60 output off-peak or $0.30 and $1.20 at peak, per million tokens. Cache hits cost $0.003 or $0.006.

In Artificial Analysis’s Ember launch comparison, DeepSeek Flash scores 33 at about $0.60 per benchmark task, against Ember’s 36 and $0.14. DeepSeek’s cheaper off-peak cache reads and lower output rate on very long prompts could still matter for a pipeline that repeatedly sends the same large body of context.

Muse Spark Open is Meta’s own route to running a million-token multimodal Muse model on infrastructure you control. Meta’s documentation lists 816 billion total parameters and 42 billion active per token. The weights come under the custom Muse Model License and an acceptable use policy.

A four-bit checkpoint still weighs in at about 537 GB before serving overhead. Together AI charges $1.25 input and $4.25 output per million tokens. Artificial Analysis lists a current score of 25 at xhigh reasoning and $2.02 per benchmark task.

The case for Spark Open is control over where the model runs and the ability to adapt it. For a team that wants to run a much smaller Muse model locally, Meta’s 30-billion-parameter Muse Glimmer is available under Apache 2.0.

Xiaomi’s MiMo-V2.6-Pro offers MIT-licensed weights, a million-token window and native text, image, audio and video input.

It has roughly one trillion parameters, with 42 billion active per token. Xiaomi’s API lists $0.435 input and $0.87 output per million, with cached input at $0.0036. Artificial Analysis currently scores it 30 at $0.26 per task.

Xiaomi’s MIT-licensed MiMo-V2.6-Flash is a closer price comparison. It scores 27 and costs about $0.13 per benchmark task, with listed rates of roughly $0.14 for input, $0.28 for output and $0.0028 for cache hits.

Meituan’s LongCat-2.5 also offers MIT-licensed weights. At 1.6 trillion parameters with roughly 48 billion active per token, it targets long coding jobs and visual tasks. Its million-token API supports up to 256,000 output tokens, and Meituan’s limited-time pricing is $0.30 uncached input, $0.006 cached input and $1.20 output per million tokens.

Meituan reports 38.5% on Terminal-Bench 4.0 using Claude Code, while Meta reports 40.9% for Ember using mini-swe-agent. The different setups make those figures a rough point of reference. A direct comparison on the same codebase and with the same tools would tell teams more about which model to use.

How Meta got from Spark to Ember

Meta Superintelligence Labs introduced the original Muse Spark in April. Following years of distributing Llama weights, Meta initially kept Spark proprietary and put it in Meta AI and a private API preview. Spark 1.1 opened up that API in a public preview in July.

In August, Spark 1.2 arrived alongside Muse Code, Meta’s terminal coding agent with persistent background jobs. Five days later, Muse Glimmer brought downloadable weights back to Meta’s model lineup.

In early September, Spark 1.3 improved coding and multi-step workflows, with its best launch scores coming from a maximum-effort configuration that was initially limited.

In October, Muse Wonder and Wonder Full created a higher tier for difficult tasks. Wonder shipped through the API and Muse Code, while Wonder Full was available to Muse Code subscribers at launch.

Days later, Spark Open extended Meta’s open-weight lineup to a larger multimodal model. On Friday, Meta began rolling out specialized models trained using the Wonder series to users of its Muse personal agent and Meta AI. Meta says Ember is the first model it trained using Wonder Full.

Where Ember fits in a production workflow

Ember can take over individual steps even if the rest of the workflow already runs on Spark. In Meta’s comparison with Spark 1.3, Ember scored 36 to Spark’s 37 at maximum effort, with benchmark costs of about $0.14 and $2.39 per task, respectively.

A practical first deployment is a step the team already knows how to grade. A coding agent can give Ember one failing test to investigate, then send the patch through continuous integration and escalate failures to a stronger model.

A support agent can use Ember to classify a ticket and fetch the relevant records while leaving account changes behind an approval step. A document pipeline can give it a batch of scanned pages and check each extracted field against the original.

Those designs let teams measure cost-per-accepted result, including retries and escalations, alongside latency and how often context pushes requests into the higher price tier.

For teams evaluating sensitive workloads, Meta says Standard-mode API content is excluded from model training. Meta has separate zero-data-retention terms for accounts with that setting enabled. Teams that need to keep inference on their own infrastructure will have to weigh the hardware and operating costs of a downloadable model against Ember’s hosted rates.

For teams that can use a hosted service, the clearest first test is a frequent, well-defined task with an existing success check and enough volume for small savings on each call to add up.