Skip to content
AG
Back to Blog

Muse Ember: Meta's model priced for the agent inner loop

November 17, 2026·5 min read·Kamai Daily·Trend·Ember

Meta's Muse Ember is a small, multimodal model priced for the boring inner loop of an agent — tool calls, routing, extraction. By Meta's own numbers it matches Claude Haiku 5.5 on the intelligence index at about a third of the cost and leads on agent benchmarks, but trails on code a human has to merge. Where it pays off, and where it still costs you.

In November 2026 Meta announced Muse Ember, which it calls its smallest and most efficient model, aimed squarely at the parts of an agent that run over and over: tool calls, routing, extraction, long-document parsing and coding subtasks. The model id is muse-ember-1.0, and you can call it today through Muse Code and the Meta Model API, as well as Google Cloud and Oracle Cloud, using the Responses API. It is served at dev.meta.ai.

If you build web apps with AI, the interesting thing about this release is not the leaderboard. It is the price tag, and where in your stack it belongs.

What Meta actually shipped

Muse Ember is a multimodal model — it natively handles text, images, audio and video — with a 1M-token context window and up to 512K tokens of output. Meta says it is the first model trained with Muse Wonder Full, and it slots into a three-model line: Muse Wonder for the hardest problems, Muse Spark for coding and agentic workflows, and Muse Ember for repeatable tasks at scale.

The pricing is the headline. Ember costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 200K tokens, rising to $0.30 / $1.50 above that. Cached input is $0.004 per million. That is a small-model rate, and Meta is explicit about the slot it is competing for: the cheap, high-volume calls where you were previously either paying a frontier model's rate or reaching for an open-weight model you had to host yourself.

How it benchmarks

Meta's framing is "intelligence per dollar," and the numbers it published are built around that. On Artificial Analysis's Intelligence Index at max reasoning, Ember scores 36 — against Muse Spark 1.3 at 37, Claude Haiku 5.5 at about 36, and GPT-6 Luna at about 28. The point is the cost to get there: Meta puts the cost per Intelligence Index task at roughly $0.14 for Ember versus $2.39 for Muse Spark 1.3, and says Ember matches Haiku 5.5's best score at about a third of the cost.

Where Ember looks genuinely strong is agent and automation work, which is exactly the job it is sold for. On AutomationBench-AA, a strict pass@1 measure of end-to-end business workflows, Meta reports 37.8% for Ember against 13.9% for Haiku 5.5 and 20.2% for Luna. On Agents' Last Exam it reports 61.6% against 57.0% and 57.1%. On Terminal-Bench 4.0, an agentic terminal-coding test, 40.9% against 39.2% and 16.4%. On the Coding Agent Index-AA, 46 against 37 and 31.

There is one clear soft spot, and to Meta's credit it is in the table: FrontierCode 2.0-AA, which judges engineering tasks on whether the change would actually merge, puts Ember at 27.2% — below Haiku 5.5's 34.9%. "Cheap and fast at agent steps" is not the same as "best at finishing a real pull request," and this is the number that says so.

On speed, two partners vouched for it. You.com measured a p95 latency of 2.2 seconds, which it described as roughly 4x to 13x faster than leading open-weight models while using 36–48% fewer output tokens. Cloudflare reported up to 30% faster "at the tail" against a comparable model it tested.

Why this matters if you build with AI

Think about where an agent actually spends its calls. A planner decides what to do a handful of times. The worker underneath it calls tools, routes between steps, extracts fields from a blob of text, and summarises documents — dozens or hundreds of times per task. That inner loop is where your token bill comes from, and it is mostly not the part that needs your most expensive model.

Ember is priced for that inner loop. At $0.10 in and $0.50 out, it is cheap enough to run on every tool call and routing decision, and by Meta's own numbers it is level with Haiku 5.5 on the intelligence index at about a third of the cost, while ahead of it on the agent and automation benchmarks. That is the "subagent" or worker-model slot in a planner/worker setup, and it is the pitch Meta is making directly. Because it speaks the Responses API on a plain muse-ember-1.0 id, it drops into existing agent code without a new SDK, and the 1M-token context plus multimodal input means it can also take the document-parsing and screenshot-reading jobs you might otherwise split across two services.

The FrontierCode result is the guardrail. If part of your pipeline writes code a human has to merge, that is the step to keep on a stronger model — Ember's weak spot is precisely the "will a maintainer accept this change?" question.

What is not known yet

Almost every headline figure here is Meta's own or Artificial Analysis's. No independent accuracy-at-volume test has been published, and the weights are private, so you cannot reproduce the numbers yourself. Meta does not disclose the parameter count, the architecture, the training data, or throughput in tokens per second — so "fast" is a partner's p95, not a number you can plan capacity against. The exact comparison model and percentile behind Cloudflare's "30% faster" claim are unstated. And the one benchmark that maps to shipping real work, FrontierCode, is where it trails. None of that makes the price less attractive; it just means the only honest way to know if Ember earns its slot in your stack is to run your own eval on your own traffic.

Amar's take

This is my opinion, not a benchmark result: the model you reach for in a subagent is not the one you reach for in the planner, and a release like this is the clearest sign yet that the market agrees. Most of what my MCP agents do all day is boring — call a tool, read the result, decide the next tool — and paying a frontier rate for those steps has always felt like buying a sports car to sit in traffic. A cheap, fast execution model is what that work wants, with the expensive model kept for the plan and for the code that has to merge.

So my plan with something like Ember is not to swap my whole pipeline over. It is to split it: measure which steps are routing and extraction versus which steps are judgment, move the first kind to the cheap model, and keep a close eye on anything that produces a diff. The number I would trust is not Meta's index — it is my own cost-per-task and pass rate on my own work, before and after. Until I have run that, this is a strong, well-priced candidate for the inner loop, and nothing more.

About the author

Amar Gupta

Amar Gupta

Senior Frontend Developer — AI & MCP

I build production frontends in React, Next.js and TypeScript — and the AI and MCP tooling behind them. 7+ years shipping web applications, from data modelling through to the deployed interface.

📍 Delhi, India · Open to Full-time

Read next