3BQRIIINO. INDEPENDENT TECH JOURNAL
← All stories

Meta / Muse / AI Models / AI News · November 10, 2026

Meta Launches Muse Ember: Its Smallest Model Yet, at $0.10 per Million Tokens

Meta released Muse Ember on 9 November 2026: its smallest and most efficient model, built for the repetitive work AI agents do all day — tool calls, coding subtasks, routing, and long-document parsing — at $0.10 per million input tokens and $0.50 per million output tokens.

Meta Launches Muse Ember: Its Smallest Model Yet, at $0.10 per Million Tokens

Meta released Muse Ember on 9 November 2026, the smallest and most efficient model in its Muse family — and the pitch is not about raw power. It is about price. Ember is built for the work AI agents repeat all day: tool calls, coding subtasks, routing, extraction, and long-document parsing. The tasks a bigger model hands off hundreds of times in a single session.

The price is the story: $0.10 per million input tokens and $0.50 per million output tokens (cached input costs $0.004 per million). Meta says Ember matches Claude Haiku 5.5's best score on the Artificial Analysis Intelligence Index at about a third of the cost, and beats OpenAI's GPT-6 Luna at every reasoning setting it plotted. At $0.14 per Artificial Analysis task, Meta claims no model on the full leaderboard scores higher for less.

What Ember is for

Agents spend most of their tokens on many small, structured steps, and Meta built Ember for exactly that layer. It can carry a business workflow end to end — pulling records, calling tools in sequence, updating systems across apps — and handle the coding subtasks an agent sends off hundreds of times a session: reading files, running tests, making targeted fixes.

It comes with a 1M-token context window and up to 512K output tokens, so it can take in a full codebase or document set in a single call. It natively understands text, images, audio, and video, so agents can read screenshots, charts, and scanned documents, or work from voice memos, meeting recordings, and video without a separate model. It can also operate a computer directly from screenshots and pick out fine visual detail.

Meta's framing of the family is blunt: "Muse Wonder plans, Muse Ember handles the volume." Wonder is for the hardest problems, Spark 1.3 for coding and agentic workflows, Ember for repeatable coding and agentic tasks at scale. Ember is also the first model Meta says was trained using Muse Wonder Full, its largest model — the frontier model teaching the small one.

The benchmarks — Meta's own, with one honest loss

Meta's benchmark table, drawn from Artificial Analysis results dated 9 November 2026 and Meta's own evaluation suite, compares Ember against Claude Haiku 5.5 and GPT-6 Luna:

One row does not go Meta's way: on FrontierCode 2.0, Haiku 5.5 leads at 34.9 against Ember's 27.2. That is the kind of detail that makes the rest of the table easier to trust — Meta published a comparison where its own model loses.

All of these are company-reported figures, drawn from Meta's chosen benchmarks and Artificial Analysis' index. They have not been independently reproduced, and "more intelligence per dollar" is Meta's framing of its own numbers. The verifiable facts are the price, the availability, and the context window.

Safety and availability

Before release, Meta evaluated Ember under its Meta Superintelligence Scaling Framework, covering chemical and biological, cybersecurity, and loss-of-control risks. Across the categories assessed, Meta says the model stayed below the framework's high-risk thresholds. Though it is a small model, Meta says it received the same core safety training and testing as its larger siblings, including tests of how it behaves when acting on a user's behalf — whether it respects user intent, handles sensitive information appropriately, resists instructions hidden in web pages, documents, or tool outputs that try to redirect it, and seeks human input before high-impact actions.

Ember is available today on the Meta Model API and Muse Code as muse-ember-1.0, with Responses API support. Muse Ember, Muse Spark, and Muse Wonder are now also available on Google Cloud and Oracle Cloud. Meta named early testing partners including Box, Cline, Oracle, Cloudflare, and You.com. Box's VP of AI Products, Yashodha Bhavnani, is quoted saying Ember "can score a contract against a company's negotiation playbook clause by clause and land on the right overall risk rating" on the Box AI Complex Work Eval.

Meta did not disclose a parameter count and did not release model weights. Meta says it has "even more efficient models in training."

Why it matters for creators and developers

Frontier models get the headlines, but the economics of AI products are decided by the small models. Every agent feature — a coding assistant that runs tests after every edit, a research tool that reads a hundred documents, a workflow that routes and extracts all day — pays the small model's price on every step. At $0.10 per million input tokens, Ember is priced to disappear into the background of a product: Meta's own arithmetic puts a full benchmark task at $0.14, against $2.39 for its own Spark 1.3.

The comparison Meta wants you to make is with Claude Haiku 5.5 and GPT-6 Luna — the other labs' small models. On Meta's numbers, Ember matches Haiku's best intelligence score at a third of the cost per task. If that holds up in independent testing, the cheapest capable agent layer just got cheaper, and the cost of shipping an agent feature falls with it.

The test, as always, is your own workload: take a repetitive agent task you already run — document extraction, test-and-fix loops, routing — and compare Ember's output and your actual bill against the small model you use today. Meta's claims are testable at $0.10 per million tokens, which is itself the point.

Sources and classification

Classification: Official Announcement (Meta, 9 November 2026). Launch facts, pricing, and benchmark scores are from Meta's announcement; capability and efficiency claims are Meta's own, with no independent verification published yet. Checked 10 November 2026.

Prefer reading in Arabic? You can translate this article into Arabic with Google Translate.