Meta Muse Ember: $0.10 per million tokens, ties Claude Haiku 5.5 at 36 on Artificial Analysis

Meta's Superintelligence Lab released Muse Ember on Monday 9 November, the smallest and cheapest model in its Muse family, aimed at the repetitive work that AI agents do all day. Meta prices it at $0.10 per million input tokens, and independent testing by Artificial Analysis puts it level with Anthropic's Claude Haiku 5.5 on its overall Intelligence Index, while finishing the benchmark tasks roughly five times faster and at about a third of the cost per task. For teams building agents, the interesting number is not the list price, it is the cost of a finished task.

Key points

• Meta released Muse Ember on 9 November 2026 as the fourth model in its Muse line, after Spark, Glimmer and Wonder (TechCrunch).
• Meta positions it as a high-volume "worker" model for tool calls, coding subtasks, routing, extraction and long-document parsing, paired with the larger Muse Wonder as planner.
• Artificial Analysis lists $0.10 per million input tokens and $0.50 per million output tokens, a 1M-token context window, and a score of 36 on its Intelligence Index (max setting).
• On that index Claude Haiku 5.5 (max) also scores 36 at the same list prices, but Artificial Analysis measures $0.41 per task for Haiku against $0.14 for Ember, and 14.8 minutes per task against 3.1.
• Ember is a proprietary model; we could not confirm how developers can call it, and Meta's own announcement page did not open for us.

What happened

According to TechCrunch, Meta Superintelligence Lab released Muse Ember on Monday as its lightest and most cost-efficient model so far. Meta describes it as built for complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction and parsing long documents.

Meta's pitch is cost per task. TechCrunch reports that Meta used Artificial Analysis's composite benchmark score to show Ember as both cheaper and smarter than Anthropic's Haiku and OpenAI's Luna, and that at 10 cents per million input tokens Ember is also cost-competitive with open models such as DeepSeek V4.1 Flash and GLM 5.3 Flash. Meta also says it trained Ember to work across text, images and audio, so that agents can read screenshots, charts and scanned documents, or work from voice memos, meeting recordings and video, without a separate model.

The intended design is a two-tier system. In the words TechCrunch quotes from Meta's announcement: "Muse Wonder plans, Muse Ember handles the volume." In a customer-support agent, for example, a heavier model could set the rules for a type of case and a lighter one apply them across tens of thousands of tickets.

Background

Ember is the fourth model released under the Muse brand, TechCrunch notes. Muse Spark, a mid-weight model, arrived in April. Its open-source twin Glimmer followed in August, and the Wonder frontier model in October. The consumer product, Muse Agent, launched in September, with features such as Spaces and video chat added earlier this month.

The cheap-worker tier is now crowded. Anthropic's Haiku line, OpenAI's Luna and and open models such as DeepSeek V4.1 Flash and GLM 5.3 Flash all compete on price for the same agent workloads, where a model is called thousands of times and small per-token savings add up. For context on the other end of the market, see our coverage of Anthropic's new Fable 5.2 and Mythos 5.2.

The numbers

Artificial Analysis, an independent benchmarking firm, already lists Ember on its model page and its Meta provider page. Its Intelligence Index v5.0.1 combines nine evaluations, including Terminal-Bench 4.0, FrontierCode 2.0, HLE-Diamond and AA-Omniscience. For Ember at its "max" reasoning setting it reports:

• Intelligence Index: 36 (34 at the "xhigh" setting), against a median of 24 for comparable models.
• Price: $0.10 per million input tokens, $0.50 per million output tokens, a 96% discount on cached input, and a blended price of $0.07 per million tokens.
• Speed: 449 output tokens per second, 4.22 seconds to the first answer token, and 3.1 minutes per Intelligence Index task.
• Cost: $0.14 per Intelligence Index task.
• Verbosity: 115 million output tokens across the index, which Artificial Analysis calls somewhat verbose (median 90 million).
• Context and modalities: 1M tokens; text and image input, text output.

The comparison with Haiku is the one Meta chose, so we checked it. Artificial Analysis's page for Claude Haiku 5.5 (max) also shows a score of 36 and the same $0.10 and $0.50 list prices, but $0.41 per task, 14.8 minutes per task and 298 million output tokens. Because Ember uses far fewer tokens to reach the same score, the identical sticker price turns into a roughly threefold difference in what a finished task costs. On this index, then, Ember ties with Haiku rather than beating it; Haiku ranks first of 33 comparable models and Ember second at the same rounded score.

Claude Haiku 5.5 (max)0.41

Muse Ember (max)0.14

Artificial Analysis cost per Intelligence Index task (US$)

	US$
Claude Haiku 5.5 (max)	0.41
Muse Ember (max)	0.14

Source: Artificial Analysis model pages, 10 November 2026

Inside Meta's own range, the index shows how the tiers differ: Muse Wonder (max) scores 44, Muse Spark 1.3 (max) 37 and Ember (max) 36. Wonder and Spark need about 12.7 and 12.4 minutes per task, against Ember's 3.1.

Claude Haiku 5.5 (max)14.8

Muse Wonder (max)12.7

Muse Spark 1.3 (max)12.4

Muse Ember (max)3.1

Time per Intelligence Index task (minutes)

	minutes
Claude Haiku 5.5 (max)	14.8
Muse Wonder (max)	12.7
Muse Spark 1.3 (max)	12.4
Muse Ember (max)	3.1

Source: Artificial Analysis (Meta provider page and Claude Haiku 5.5 page)

What it means

For developers, the useful reading is that token price is a poor guide for agent workloads. A model that thinks less per task can cost far less even at the same list price, and finish sooner, which matters when an agent chains dozens of calls. If Artificial Analysis's numbers hold on your own tasks, Ember is a candidate for the high-volume steps of a pipeline while a stronger model plans.

For Meta, Ember fills the bottom of a lineup that now runs from a frontier model to a cheap worker. It also shows how benchmark framing works: Meta picked Artificial Analysis's index and the cost-per-task view, which suits a model that is frugal with tokens. The independent data supports the cost and speed claims, but not "smarter" than Haiku on the overall score.

Not every workload will favour it. Artificial Analysis's index leans toward agentic and coding tasks, and results on your own prompts can differ, so test before switching.

What we don't know yet

• How developers get access. A chart on Artificial Analysis's Ember page carries the note "No publicly available API", although the site lists prices; we could not confirm availability, rate limits or which platforms offer it.
• Whether audio input is supported. Meta, via TechCrunch, says Ember handles text, images and audio; Artificial Analysis lists text and image input and text output.
• Meta's own benchmark details. We could not open Meta's announcement, so we have only TechCrunch's account of what it says.
• Whether the model's weights will ever be released. Artificial Analysis lists Ember as proprietary; Glimmer is the open-source sibling of Spark, and no open Ember has been announced in our sources.

What to watch next

Watch for Meta's API documentation and pricing page, independent results on agent benchmarks beyond Artificial Analysis, and whether Anthropic, OpenAI or DeepSeek respond on price or token efficiency. Also watch whether the benchmark gap holds as Artificial Analysis refreshes its measurements, which it repeats every six hours for current models.

Sources

• TechCrunch: Meta releases Muse Ember, a new lightweight model for agent execution
• Artificial Analysis: Muse Ember (max)
• Artificial Analysis: Meta provider page
• Artificial Analysis: Claude Haiku 5.5 (max)