Meta Muse Ember: Cheap Agent Model Explained

This guide covers what Meta announced, how the numbers compare with Claude Haiku 5.5 and GPT-6 Luna, where independent testers agree and where they do not, and how a small model fits into a planner-and-worker agent design. Most numbers are Meta's own, and several come from internal benchmarks, so we flag which are independent.

• What it is: Meta's smallest and most efficient Muse model, model ID `muse-ember-1.0`, announced 09/11/2026. It is the first model trained using Muse Wonder Full.
• Price: $0.10 input, $0.004 cached input and $0.50 output per million tokens for prompts up to 200K tokens. Longer prompts cost $0.30, $0.012 and $1.50.
• Specs: 1M-token context, up to 512K output tokens, native text, image, audio and video input, and computer use from screenshots. Parameter count and weights are undisclosed.
• Independent score: Artificial Analysis Intelligence Index 36, tied with Haiku 5.5 at max reasoning and ahead of GPT-6 Luna at 28.
• Where it wins: agent automation (AutomationBench-AA), computer use (OSWorld 2.1) and multimodal perception, per Meta's table.
• Where it loses: FrontierCode 2.0-AA, where Haiku 5.5 scores 34.9 against Ember's 27.2.
• Availability: Muse Code, the Meta Model API, Google Cloud and Oracle Cloud.

Meta positions three tiers. Muse Wonder handles the hardest problems, Muse Spark covers coding and agentic workflows, and Muse Ember handles repeatable coding and agentic tasks at scale. Our Muse Wonder review covers the top tier and the cost caveats, and our Muse Spark 1.3 review covers the middle tier, which Artificial Analysis lists at $2.39 per task and an Intelligence Index of 37.

That comparison is the heart of the Ember story. According to Meta's chart, Ember scores 36 on the same index at about $0.14 per task, against 37 for Spark 1.3 at $2.39. The index points are close, and the cost is about seventeen times lower. If you accept the benchmark as a fair proxy for your work, the premium for the larger model buys a one-point gap. Our own view is that a single composite score hides a lot, so test on your own tasks.

An agent that plans, calls tools, reads results and retries can make dozens or hundreds of model calls per task. If every call goes to a frontier model, cost and latency add up. The standard fix is a division of labour: a strong planner sets the approach, a cheap fast worker does the volume, and a reviewer checks the output. Meta describes exactly this pairing, with Muse Wonder planning on the hardest problems and Ember handling the volume. Julian Goldie's video gives it a plain framing: a planner sets the rules, a fast worker does the volume and a reviewer checks the work.

This is the same pattern you see in open agent frameworks, which we explain in our Hermes Agent guide, and in Meta's own agent work in Muse Agent in a secure VM. The same day's flagship launch, Claude Fable 5.2, makes the planner side of the same bargain: a stronger and now cheaper top-end model that you would call sparingly while smaller models do the volume.

Meta lists the intended jobs plainly:

• Business workflows end to end: pulling records, calling tools in sequence and updating systems across apps.
• Coding subtasks: reading files, running tests and making targeted fixes.
• Ingesting a full codebase or document set in one call, using the 1M-token window.
• High-throughput agent and subagent workloads.
• Extraction, routing and long document parsing.

Sterling is approximate at about £0.79 to the dollar. The list price of Ember matches GPT-6 Luna exactly, so price alone does not separate them. What separates them is how many tokens each model burns and how well it performs. Artificial Analysis cost-per-task figures in Meta's post show Ember at $0.116 for a score of 33.9 and $0.136 for 35.8 at maximum, against Haiku 5.5 at $0.092 for 24.6 rising to $0.411 for 36.2 at maximum. Meta's reading is that Ember matches Haiku's best score for about a third of the cost, and beats every GPT-6 Luna setting. For Luna, see our GPT-6 Sol and Luna guide.

Cached input is where the saving gets dramatic: $0.004 per million cached tokens means that repeating a long system prompt or a large document costs almost nothing. For agent loops that resend the same context, that line item matters more than the headline price.

Meta reports these numbers at maximum reasoning effort for all three models. Some are from Artificial Analysis, which is independent, and some are Meta internal evaluations, marked below.

How to read the table

Three points matter. First, the agentic gap is large on AutomationBench-AA, where Ember scores nearly three times Haiku's result, which is why Meta leads with it. Second, Terminal-Bench 4.0 compares different harnesses: Ember with mini-swe-agent, Haiku with Claude Code and Luna with Codex. That is described as the best published harness for each, but harness choice can move a score by several points, so a 1.7-point lead over Haiku is within what setup differences can produce. Third, Haiku wins FrontierCode 2.0-AA by 7.7 points, which is the most developer-relevant gap in the table. Coverage from outlets reading Meta's figures reached the same conclusion: if your workload is hard software engineering, Haiku 5.5 still looks better.

Secondary coverage also quotes different AutomationBench figures from the 37.8% strict pass@1 in Meta's post, which probably reflects different benchmark variants, so quote only numbers from the primary source and always name the variant. For our review of the rival small model, see Claude Haiku 5.5.

Independent checks

Plotly ran Ember on its data analytics benchmark and reported that it passed 38 of 43 questions (88%) at a cost of $1.04, tying GPT-6 Luna and beating Haiku 5.5 by a point. A single small benchmark is a weak basis for conclusions, but it points the same way as the Artificial Analysis cost-per-task result: Ember is competitive at a low cost for structured tasks.

Multimodal input

Ember accepts text, images, audio and video natively. Meta says it can read screenshots, charts, scanned documents, voice memos, meeting recordings and video, and can operate a computer from screenshots while picking out fine visual detail. A small model that handles audio and video natively is valuable for pipelines that otherwise need separate transcription steps, such as processing meeting recordings or scanned invoices.

Long context at low cost

A 1M-token window and 512K maximum output allow whole-repository or document-set ingestion in one call. The long-prompt rate triples the price, but at $0.30 per million input tokens it is still cheaper than most models' base rate. As always with long context, check retrieval quality on your own data. A large window does not guarantee the model uses every part of it well.

Speed

Partner feedback in Meta's post is useful, if selective. Cloudflare's CTO said the model was up to 30% faster at the tail than the comparable model tested, You.com reported a p95 latency of 2.2 seconds, about four to thirteen times faster than leading open-weight models, with 36% to 48% fewer output tokens, and Box praised accuracy, speed and token efficiency on clause-by-clause contract scoring and candidate screening. Cline's founder described it as suited to subagent work. These are partner quotes selected by Meta, but the pattern, lower latency and fewer tokens, is consistent.

Meta evaluated Ember under its Superintelligence Scaling Framework, covering chemical and biological risk, cybersecurity and loss of control, and found it below the framework's high-risk thresholds in all assessed categories. It received the same core safety training and testing as the larger models. The agent-behaviour evaluation covers respecting user intent, handling sensitive information, resisting hidden instructions in web pages, documents or tool outputs, and seeking human input before high-impact actions. Prompt injection is the most relevant risk for a model that reads untrusted documents all day, so test it yourself and keep humans in the loop for actions that change data.

The architecture Meta describes is easy to state and easy to get wrong in practice. Here is a pragmatic way to adopt it without betting production traffic on a new model.

Step 1: Classify your calls

Log a week of agent traffic and sort the calls into three buckets: planning (deciding what to do), execution (calling a tool, extracting a field, editing a file) and review (checking an output). In most agent systems execution calls outnumber planning calls by a wide margin, which is where a model priced at $0.10 per million input tokens makes the largest dent in the bill. Planning and review calls are fewer, and they are where a stronger model such as Fable 5.2 or Muse Wonder earns its price.

Step 2: Run a shadow test

Send the same execution calls to Ember and to your current model, then grade the outputs against the cases you already know are right. Pay particular attention to tool-call formatting, because a small model that returns malformed arguments one time in fifty can cost more in retries than it saves in tokens. Because Ember's FrontierCode 2.0-AA score trails Haiku 5.5, include a handful of genuine coding edits in the sample, not only extraction and routing.

Step 3: Exploit caching

At $0.004 per million cached input tokens, the cheapest design is one that keeps a long, stable prefix, such as tool definitions, policies and reference documents, and varies only the tail. A prompt that changes at the front defeats caching and removes much of the advantage. Structure prompts so the stable material comes first.

Step 4: Keep a reviewer

A cheap worker needs a check. That can be a stronger model reviewing a sample of outputs, a rule-based validator or a human approval step before high-impact actions. Meta's own safety notes stress seeking human input before high-impact actions, and that is sensible advice for any pipeline that lets a small model change real data.

Step 5: Watch the long-prompt tier

Prompts over 200K tokens cost three times the base rate. A workflow that routinely sends 400K-token prompts is paying the higher tier on every call, so it can be cheaper to chunk documents or retrieve only relevant sections than to rely on the full 1M window.

Benchmarks for small models are especially easy to misread, for three reasons. First, composite indices average very different tasks. Ember and Haiku both score 36 on the Intelligence Index, yet Meta's own table shows Ember ahead by roughly three times on one agent benchmark and behind by nearly eight points on another. The same total can hide opposite strengths, so the index is a guide to general standing, not a recommendation for any one job.

Second, cost per task depends on token use as well as price. A model that reasons for longer at max effort burns more output tokens, and the Artificial Analysis cost curves in Meta's post show how Haiku 5.5's cost rises from $0.092 to $0.411 as its reasoning effort rises, while its score climbs from 24.6 to 36.2. Ember's two plotted points sit at a narrower cost range. That makes budgeting easier, but it also means you have fewer settings to tune.

Third, partner quotes show what vendors want shown. The Cloudflare, You.com, Box, Cline and Oracle comments in Meta's post are positive by construction. They are still informative, because they name specific workloads (multi-repo investigation, contract scoring, subagent coding) that match the model's intended role. What they cannot tell you is how it performs on your data, so the shadow test above matters more than any quote.

The one independent data point we have beyond Artificial Analysis is Plotly's data analysis benchmark, where Ember's 38 of 43 correct answers for $1.04 tied GPT-6 Luna. The sample is small, and it covers structured analysis, not open-ended coding, so it supports the narrow claim that Ember is competent at cheap structured work and no more.

• Against Claude Haiku 5.5: Ember matches the index score at roughly a third of the cost per task and wins on agent automation and computer use, but trails on FrontierCode 2.0-AA.
• Against GPT-6 Luna: the list prices are the same, and Ember scores higher on every benchmark Meta shows except one tie on GDP.pdf-AA.
• Against Muse Spark 1.3: nearly the same index score at a small fraction of the cost per task, although Spark is the better pick for harder coding.
• Against Fable 5.2: not a like-for-like contest. Fable 5.2 costs 80 times more per input token and leads on hard reasoning and long-horizon work. See our Fable 5.2 review.

• Weights and parameter count are not published, so you cannot self-host or inspect it.
• Several of the strongest results are Meta internal benchmarks that outsiders cannot reproduce.
• Harnesses differ across models in the Terminal-Bench comparison.
• Haiku 5.5 still leads on hard software engineering tasks.
• Cost per task depends on how many tokens a model uses, so your real costs will differ from the benchmark's.
• It is new: expect independent testing to shift the picture over the coming weeks.

• Teams running high-volume agents with repetitive tool calls, routing or extraction.
• Document-heavy workflows such as contract review, invoice processing and meeting-recording analysis.
• Multi-model architectures that pair a premium planner with cheap workers.
• Not ideal for: hard debugging, large refactors or tasks where a single error is expensive. Use a stronger model there.

A practical approach: route a sample of your real workload to Ember and a stronger model, grade both, and move over only the task types where the cheaper model matches.

Muse Ember is the most interesting cheap model of the week: tied with Haiku 5.5 on an independent index, ahead on agent automation and computer use in Meta's tables, and priced the same as GPT-6 Luna with better scores. Its weak spot is hard coding, and many of the best numbers come from Meta itself. Treat it as a strong candidate for the worker role in an agent system, and verify it on your own tasks before moving production traffic.

• Meta: Introducing Our Smallest, Most Efficient Model, Muse Ember (09/11/2026): price, specs, benchmarks and safety.
• Plotly: Meta Muse Ember benchmark results for data analysis: independent data analysis test.
• Digital Trends, MarkTechPost and Startup Fortune: launch coverage.
• Julian Goldie SEO on YouTube: creator explainer.
• Image: header artwork from Meta's announcement, credited to Meta and compressed to AVIF.