Meta's Muse Ember matches Claude Haiku 5.5 at a third of the cost per task

What's happening

Meta Superintelligence Labs, Meta's AI division led by Alexandr Wang, unveiled Muse Ember on Monday, November 9, 2026. Meta calls it its smallest and most efficient AI model so far. It's built for the work AI agents repeat all day long. AI agents are programs that complete tasks on their own across multiple steps. Meta's list of that work includes tool calls, coding subtasks, routing requests, pulling information out of text and analyzing long documents.

Ember is available now in Muse Code, Meta's coding environment, and through the Meta Model API. The model ID is muse-ember-1.0, and it also supports the Responses API, a common interface for building agent applications. Ember isn't an open model, though. As with earlier Muse models, Meta is keeping the weights under wraps, according to shiftdelete.net.

Same price list, fewer tokens per task

Here's what Meta charges:

Tokens are the chunks of text a model reads and writes, and they're what you get billed for. At 10 cents per million input tokens, TechCrunch calls Ember "significantly cheaper" than Anthropic's Haiku and OpenAI's Luna, and puts it on par with open models like DeepSeek V4.1 Flash and GLM 5.3 Flash. Other sources disagree on the first point. According to The New Stack (quoted by shiftdelete.net) and OfficeChai, the list prices are identical to GPT-6 Luna's ($0.10/$0.50). They also match Claude Haiku 5.5's standard rate for inputs up to 100,000 tokens.

So Meta's real pitch is the cost per completed task. That depends on how many tokens a model burns before it's done. A model that "thinks" less and answers more concisely costs less at the same rate. For agents firing off thousands of requests a day, that's the number that ends up on the bill.

According to shiftdelete.net, though, teams will likely save the most through the very low cache price. Inputs that never change, like long instructions or documents, get cached and billed at a fraction of the normal rate. The context window is 1 million tokens, meaning the amount of text the model can take into account at once. Meta says that's enough to fit an entire codebase or document collection into a single request. Ember can output up to 512,000 tokens in one response.

Cryptobriefing points out something else: the prices sit close to the Meta Model API's discounted "Contributor" tier. If you use that tier, you allow Meta to use your interaction data to train future models. For companies with strict privacy requirements, that could be a dealbreaker.

The benchmarks in detail

At its maximum reasoning level, Ember scores 35.8 points on the Artificial Analysis Intelligence Index. That's a composite score from independent analytics firm Artificial Analysis that rolls many tests into a single number. Claude Haiku 5.5 scores 36.2 points, practically a tie. But according to OfficeChai, a task costs about $0.136 with Ember versus $0.411 with Haiku. GPT-6 Luna manages only 27.8 points at its maximum level, for $0.152. Even Ember's lower "xhigh" level (33.9 points for $0.116) beats Luna's best result. Meta also says no model on the entire leaderboard scores more points for less money per task.

Individual results at maximum reasoning level, all according to Meta:

Ember's biggest lead is on AutomationBench-AA, which tests complete business workflows end to end. There it scores almost three times as high as Haiku. According to Meta, Ember can retrieve records, chain tool calls and update systems across multiple applications. OSWorld measures computer use, meaning clicking and typing on a real desktop.

The internal comparison is just as stark. Muse Spark 1.3 scores 37 points on the index but costs $2.39 per task (token prices $1.25/$4.25). So for one point less, Ember is about 17 times cheaper.

Planner and worker: how Ember is meant to be used

Meta sees Ember as a subagent: a small model that works through tasks handed out by a larger planning model. Meta's shorthand for it:

"Muse Wonder plans, Muse Ember does the bulk of the work."

TechCrunch gives a customer service example. A heavyweight model decides how a certain type of problem should be handled, and Ember applies those rules to tens of thousands of cases. Muse Wonder costs $4 per million input tokens and $20 per million output tokens. According to OfficeChai, it's the same planner-worker pattern Anthropic recommends for Haiku 5.5 alongside Opus 5.5 and Sonnet 5.5. The roles in Meta's model family: Wonder for the hardest problems, Spark for coding and agent workflows, Ember for repeatable high-volume tasks.

Ember is natively multimodal: it processes text, images, audio and video directly, without add-on models. That means an agent can analyze screenshots, charts, scanned documents, voice memos, meeting recordings or videos. Ember is also the first model Meta trained with help from Muse Wonder Full, the large model it launched in October.

Early testers and safety

Early testers include Box, Cline, Oracle, Cloudflare and You.com, according to shiftdelete.net. Cloudflare CTO Dane Knecht says Ember ran "up to 30% faster at the slow end" than the comparable model his team tested. That refers to the slowest responses, which are often the most disruptive in practice. He says it also performed well on investigations spanning multiple code repositories. Chak Pothina, head of inference at You.com, cites a p95 latency of 2.2 seconds, meaning 95 percent of responses arrived faster than that. He says Ember generated 36 to 48 percent fewer output tokens than the models it was compared against.

On safety, Meta says Ember was evaluated under its in-house Superintelligence Scaling Framework. It stayed below the high-risk thresholds for chemical and biological risks, cybersecurity and loss of control. Meta also tested its resistance to prompt injection: hidden instructions in web pages, documents or tool outputs designed to hijack an agent. And Meta checked whether Ember asks a human before taking consequential actions.

Background: Meta's Muse family

Ember is the fourth model under the Muse brand. Before it came the mid-sized Spark in April, its open-source twin Glimmer in August and the large flagship Wonder in October. In September, Meta also launched Muse Agent, a consumer assistant that TechCrunch says has been a surprise hit. It has since gained features like Spaces and video chat, and partners such as Shopify and Expedia have signed on. According to OfficeChai, a model cheap enough for every product fits that strategy. Meta says even more efficient models are already in training.

Our take

With Ember, Meta is going after the spot of cheap workhorse model in agent systems, currently held by Haiku and Luna. If you run agents at scale, low cost per task, a very cheap cache and a million-token context are solid arguments.

There are some caveats, though. Meta measured several benchmarks itself, including Agents' Last Exam, OSWorld 2.1, PerceptionBench and its in-house Personal Superintelligence Index. On Terminal-Bench, each model ran in a different test harness: Ember with mini-swe-agent, Haiku with Claude Code, Luna with Codex. So it's not an apples-to-apples comparison. The index scores come from the new version 5, which relies mostly on non-public datasets. That's why Haiku scores only 36 there instead of its earlier 43, and old and new scores can't be compared. Ember's costs are also based on Meta's list prices. On real coding work like FrontierCode, Haiku is well ahead. shiftdelete.net therefore recommends testing Ember on your own tasks before switching. And if openness matters to you, there's nothing to download here: Ember is only available through Meta's API.