Meta’s Muse Ember matches Haiku at about a third the cost

Muse Ember, Meta’s smallest Muse model, went live on November 9 in Muse Code and the Meta Model API as muse-ember-1.0. If your agents run subtasks on Claude Haiku 5.5 or GPT-6 Luna, Ember charges the same $0.10 per million input and $0.50 per million output tokens, and Meta’s announcement puts it at about $0.14 per Artificial Analysis Intelligence Index task at max reasoning, against $0.41 for Haiku. Meta labels those Ember costs as estimates at proposed pricing, and Haiku still leads on FrontierCode 2.0-AA, 34.9% to 27.2%.

Meta built Ember for the work agents repeat all day: tool calls, coding subtasks, routing, extraction and long-document parsing. It has a 1M-token context window, returns up to 512K output tokens and supports the Responses API.

Price per task

The New Stack notes that Ember’s input and output rates match Luna’s and Haiku’s standard short-context prices. Both of those models charge $0.01 per million cached input tokens at those rates, with higher prices for longer prompts. Ember’s cached input costs $0.004 per million.

Meta’s chart plots each reasoning-effort setting against cost per task, using raw Artificial Analysis results. At max, Ember scores 35.8 for $0.136 per task, and Haiku at max scores 36.2 for $0.411. Both round to 36, so Meta’s claim that Ember “matches Haiku’s best score” holds only after rounding. One setting down, at xhigh, Ember costs less than half as much as Haiku: 33.9 for $0.116, against 34.3 for $0.248. Luna at max scores 27.8 for $0.152, below Ember at xhigh for more money.

Ember still isn’t the cheapest option per task. Haiku at medium costs $0.092 and Luna at high costs $0.073, but they score 24.6 and 21.4.

Who measured what?

All the benchmark scores Meta published are at max reasoning. Meta’s charts credit four of them to Artificial Analysis:

• AutomationBench-AA (end-to-end business workflows, strict pass@1): Ember 37.8%, Luna 20.2%, Haiku 13.9%.
• Coding Agent Index-AA: Ember 46, Haiku 37, Luna 31.
• FrontierCode 2.0-AA, which judges real engineering tasks on whether the result could be merged: Haiku 34.9%, Ember 27.2%, Luna 25.2%.
• GDP.pdf-AA (professional work over long documents): Ember and Luna tied at 22.8%, Haiku 20.8%.

Haiku’s FrontierCode lead matters because coding is where Meta pitches Ember hardest: the subtasks an agent sends out hundreds of times a session, such as reading files, running tests and making targeted fixes. For the hardest problems, Meta suggests pairing it with Muse Wonder, which plans while Ember “handles the volume.”

Meta labels three results as its own internal evaluations. On Agents’ Last Exam, run on the Linux subset, Ember scores 61.6% against 57.1% for Luna and 57.0% for Haiku. OSWorld 2.1 tests computer use across desktop apps: with partial credit, Ember scores 69.5% to Haiku’s 68.9%, and with binary scoring, where a task either succeeds or fails, 33.8% to 27.8%, with Luna at 17.0%. On PerceptionBench with tools, Ember scores 80.6%, Haiku 78.6% and Luna 62.8%. Meta’s Personal Superintelligence Index, an internal composite of personal and knowledge work, has Ember at 46.8% against 37.0% for Luna and 36.1% for Haiku. On Terminal-Bench 4.0, where each model runs in its best published harness, Ember scores 40.9%, Haiku 39.2% and Luna 16.4%.

Against the Chinese Flash models

The New Stack set Ember against three Chinese models: DeepSeek V4.1 Flash 1017, Z.ai’s GLM-5.3 Flash and MiniMax M3.1 Flash. Ember’s input price is lower than the listed uncached input rate of all three. Its output price is lower than DeepSeek’s direct off-peak rate and MiniMax’s listed rate, and the same as GLM’s. DeepSeek is slightly cheaper on one line, its $0.003 off-peak rate for cached input. On the Artificial Analysis Intelligence Index, the outlet puts DeepSeek at 33, MiniMax at 29 and GLM at 28, against Ember’s 36. For reference, M3.1 Flash launched in October at $0.15 input and $0.60 output per million tokens.

Ember in Meta’s lineup

Ember is the first model Meta trained using Muse Wonder Full. It takes text, images, audio and video, so an agent can read screenshots, charts, scanned documents or meeting recordings without a separate model. Meta compares it with Muse Spark 1.3, which it pitches for coding and agentic workflows: at max reasoning, Spark 1.3 scores 37 on the Artificial Analysis index, one point above Ember, but costs $2.39 per task and $1.25/$4.25 per million input/output tokens.

As with earlier Muse models, Meta is keeping Ember’s weights private, The New Stack reports. The outlet expects Ember to reach OpenRouter and the neoclouds soon, because Meta has done that with its other Muse models.

Meta evaluated Ember under its own Meta Superintelligence Scaling Framework, covering chemical and biological, cybersecurity and loss-of-control risks, and says it stayed below the framework’s high-risk thresholds. A separate evaluation looked at how Ember behaves when acting for a user, including whether it resists instructions hidden in web pages, documents or tool outputs and asks a human before high-impact actions.

Box tested Ember before launch. Yashodha Bhavnani, its VP of AI Products, said it can “score a contract against a company’s negotiation playbook clause by clause and land on the right overall risk rating.”