While Anthropic’s Fable 5.2 grabbed the frontier headlines on November 9, Meta shipped a model aimed at the opposite end of the market. Muse Ember is the smallest and most efficient model from Meta Superintelligence Labs, priced at $0.10 per million input tokens and built for “the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing.” Independent benchmarker Artificial Analysis says it matches Claude Haiku 5.5’s top score at about a third of the cost per task.
Quick Take
- What it is: A low-cost, multimodal reasoning model (
muse-ember-1.0) on the Meta Model API and in Muse Code, with a 1M-token context window and up to 512K output tokens. - Price: $0.10 input, $0.004 cached input and $0.50 output per million tokens up to 200K-token prompts; $0.30/$0.012/$1.50 above that.
- Performance: 36 on the Artificial Analysis Intelligence Index at max effort, the same rounded score as Claude Haiku 5.5, for about $0.14 per task versus $0.41.
- Weak spot: On FrontierCode 2.0-AA, which judges real engineering changes on mergeability, Haiku 5.5 leads 34.9% to 27.2%.
- Strategy: “Muse Wonder plans, Muse Ember handles the volume” — Ember is the worker model in a two-tier agent setup.
What happened
Meta announced Ember in a research blog post, claiming it “delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting.” Meta’s Alexandr Wang wrote that Ember “offers nearly Muse Spark 1.3 (max) intelligence at a 17x cheaper price point (per AA)” and is “the first model we’ve released trained using Muse Wonder Full.” It launched with partner quotes from Box, Cline, Oracle, Cloudflare and You.com, and Vercel made it available on its AI Gateway the same day.
Artificial Analysis published a launch evaluation calling Ember (max) “the cheapest model scoring 36 or higher, at $0.14 per task,” and noting that it is about 4x cheaper than DeepSeek V4.1 Flash 1017 (score 33, $0.60) and GLM-5.3-Flash (28, $0.58).
Background: Meta’s Muse lineup
Ember is the fourth model under the Muse brand, as TechCrunch notes: the mid-weight Spark (April), its open-source twin Glimmer (August), and the frontier Wonder model (October). The consumer Muse agent launched in September. Meta now describes a three-part portfolio: “Muse Wonder for the hardest problems, Muse Spark for coding and agentic workflows, and Muse Ember for repeatable coding and agentic tasks at scale.” On Artificial Analysis, Ember (36) sits just below Muse Spark 1.3 (37) but costs $0.14 per index task against $2.39. Earlier i10x coverage of Meta’s agent push: Meta Muse AI agent launch.
How it works
Meta pitches Ember as the execution layer of an agent system. A larger model sets the plan or the rules; Ember applies them across many cases. In TechCrunch’s customer-support example, a heavier model decides how to handle an issue, then a lighter model applies those rules across tens of thousands of tickets. Ember natively accepts text, images, audio, video and PDFs, so “agents can read screenshots, charts, and scanned documents, or work from voice memos, meeting recordings, and video without a separate model.” According to VentureBeat, it is served through Responses, Chat Completions and Messages endpoints, with reasoning effort from minimal to max (max is the default), and is also listed via Oracle Cloud and in preview on Google Cloud.
Much of the per-task advantage comes from caching. Artificial Analysis notes that Ember has the same input and output list prices as Haiku 5.5 and GPT-6 Luna, but its cached input — “most of the cost in our agentic evaluations” — is 60% cheaper.
| Benchmark (max effort) | Muse Ember | Claude Haiku 5.5 | GPT-6 Luna | Source |
|---|---|---|---|---|
| Artificial Analysis Intelligence Index | 36 (35.8) | 36 (36.2) | 28 | Artificial Analysis |
| Cost per AA index task | $0.14 | $0.41 | $0.15 | Artificial Analysis / Meta |
| AutomationBench-AA | 37.8% | 13.9% | 20.2% | Artificial Analysis |
| Terminal-Bench 4.0 | 40.9% | 39.2% | 16.4% | best published harness |
| Coding Agent Index-AA | 46 | 37 | 31 | Artificial Analysis |
| FrontierCode 2.0-AA | 27.2% | 34.9% | 25.2% | Artificial Analysis |
| OSWorld 2.1 (partial credit) | 69.5% | 68.9% | 50.5% | Meta internal |
| Agents’ Last Exam | 61.6% | 57.0% | 57.1% | Meta internal |
Meta’s chart shows the unrounded index scores as 35.8 for Ember and 36.2 for Haiku 5.5, so “matches” means “ties after rounding.” Several rows are Meta’s own evaluations, not Artificial Analysis runs.
Safety
Meta says it evaluated Ember under its Meta Superintelligence Scaling Framework for chemical and biological, cybersecurity and loss-of-control risks and found it “remained below the framework’s high-risk thresholds.” It also tested whether the model respects user intent, handles sensitive information, resists instructions hidden in web pages or tool outputs, and “seeks human input before high-impact actions.” Meta did not publish detailed results in the blog post.
Pros and cons
For users
- Pro: Cheaper models make it economical to put AI into everyday products; Meta says the cost is “low enough to run in every product, for everyone.”
- Con: Users will rarely know which model handles their request, and small models trade away depth — Grok’s own assessment is that Ember “trails on deep knowledge.”
For developers
- Pro: Drop-in API endpoints, a 1M-token window, multimodal input and leading scores on agentic benchmarks like AutomationBench-AA make it a strong subagent candidate; Cline’s founder called it “purpose-built for, and excels at, subagent work.”
- Con: It is proprietary, prices triple above 200K-token prompts, and it trails Haiku 5.5 on the mergeability-focused FrontierCode benchmark.
For companies
- Pro: Box reports that Ember can score contracts “clause by clause” against a negotiation playbook; Cloudflare saw it run “up to 30% faster at the tail” than a comparable model.
- Con: Building on Meta’s API adds vendor dependency on a platform whose enterprise business is only weeks old, and cost estimates at launch were at “proposed pricing.”
What X is saying
today we released our smallest model yet: Muse Ember 🔥 Muse Ember offers nearly Muse Spark 1.3 (max) intelligence at a 17x cheaper price point (per AA). Muse Ember is the first model we’ve released trained using Muse Wonder Full. excited for y’all to try on API and muse code!
— Alexandr Wang (@alexandr_wang) November 9, 2026 · View on X
Meta’s newest model Muse Ember scores 36 on the Artificial Analysis Intelligence Index, matching Claude Haiku 5.5 (max) at a third of the Cost per Task
— Artificial Analysis (@ArtificialAnlys) November 9, 2026 · View on X
Pi Dash + Muse Ember == Coding beast! 🔥 Pi Dash is honored to get early access to Muse Ember over the past weekend!
— Pi Dash (@pidashai) November 9, 2026 · View on X
Developers mostly reacted to the price. Chinese AI commentator @dotey summarized it as the same unit price as Claude Haiku 5.5 but one-third of the cost on the same test suite. Workflow startup ERP.AI highlighted that “Muse Ember scored 37.8 on AutomationBench. The next best scored 20.2.” Meta researcher Wenhu Chen asked developers to “test it and provide feedback to us!”
What Grok says
Asked whether Ember is good for personal tasks, Grok answered: “Yes, solid for personal tasks. Muse Ember is built for high-volume everyday work: quick tool calls, summaries, extraction, light coding subtasks, and multimodal parsing at very low cost. It matches Haiku-level scores cheaply but trails on deep knowledge and some complex agentic…” In another reply, Grok noted that on the Artificial Analysis index Ember “scores 36 vs Grok 4.7’s 35.”
Critical analysis
Ember’s real story is economics, not intelligence. Small models have been “good enough” for routing and extraction for a while; what Meta claims is a better cost per finished task, driven largely by cheaper cached input. That matters because agent workloads resend context constantly — the bill is dominated by cache reads, not headline token prices.
The benchmark picture is less one-sided than the marketing. Ember wins big on AutomationBench-AA and Coding Agent Index-AA, but loses clearly on FrontierCode 2.0-AA, the test closest to “would a maintainer merge this?” That fits Meta’s positioning — Ember as a worker, not an engineer — but teams should not read “matches Haiku” as “replaces Haiku everywhere.” Several headline results (Agents’ Last Exam, OSWorld, PerceptionBench, the internal Personal Superintelligence Index) are Meta’s own runs.
Strategically, Ember is a distribution play. Meta is building an enterprise business from scratch, and a near-free worker model is an easy way into developers’ agent harnesses — the same harnesses currently routing subagents to Haiku.
What’s still open
- Will Artificial Analysis and others reproduce the agent and coding results at scale?
- Will Meta release open weights for an Ember-class model, as it did with Glimmer?
- How will Anthropic and OpenAI respond on small-model pricing, especially cached input?
- What do the “even more efficient models in training” that Meta teased look like?
About the analysis
Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated illustration. Read our AI Use Policy. Spotted an error? Request a correction.
This independent, research-based i10x analysis draws on Meta’s November 9 blog post, Artificial Analysis’s launch evaluation, reporting by TechCrunch, The Verge and VentureBeat, and public posts on X through November 10, 2026. i10x did not test Muse Ember directly.
i10x Perspective
Agents are turning AI pricing into a two-tier market: a few expensive “planner” models and a flood of cheap “worker” calls. Muse Ember is Meta’s bid to own the worker tier, and its cached-input pricing is aimed squarely at the way agent harnesses actually spend money. For builders, the lesson from Anthropic’s and Meta’s launches on the same day is the same: pick models per step, measure cost per finished task, and keep the judgment calls on the strongest model you can afford.
