# Meta's New AI Costs 10 Cents — and That's Why Your Next App Will Have an Agent in It > Source: > Published: 2026-11-10 04:42:21+00:00 By Shakeel Ahmed — AI Growth Lab | November 10, 2026 *Originally published on [AI Growth Lab](https://aigrowthlabshakeel.blogspot.com/2026/11/claude-fable-mythos-split-access.html)* On November 9, 2026, Meta released Muse Ember, its smallest AI model. It costs $0.10 per million input tokens and $0.50 per million output tokens, with a 1-million-token context window. Meta built it for the repetitive work AI agents do all day: tool calls, routing, extraction, coding subtasks, and reading long documents. That price is the whole story. The moment an AI agent costs almost nothing to run, every app starts hiring one. This article explains what Muse Ember is, what 10 cents actually buys, why cheap agents change who can build AI products, and where the small model still falls short. Meta's Muse family is the company's line of AI models for coding and agents, built by Meta Superintelligence Labs. Muse Ember is the smallest and cheapest member of that family. It went live on November 9 as `muse-ember-1.0` on the Meta Model API and inside Muse Code, Meta's coding agent. It is also listed on OpenRouter. Here is the short answer to what it does: **Ember is built for the high-volume steps that make up most of an agent's workload.** A modern AI agent does not make one big decision. It makes hundreds of small ones — call a tool, read the result, decide the next step, extract a fact from a document, route a request to the right place, repeat. Each of those steps is a separate call to a model, and each call costs money. Ember exists to make those calls cheap. Meta's launch post describes it plainly: Ember is designed "for complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing." It accepts text, images, audio, video, and PDFs as input and returns text. It can take in a million tokens at once — roughly 750,000 words — and write up to 512,000 tokens in reply. Meta CEO Mark Zuckerberg put it this way in a post on X: Ember is "our smallest model that delivers higher intelligence per dollar than others in its class." That phrase — intelligence per dollar — is the metric that matters here, and it is worth understanding properly. The headline price is $0.10 per million input tokens and $0.50 per million output tokens. But the real math is more interesting than the headline, because there are three prices, not one. | What you send | Price per million tokens | |---|---| | Input, up to 200,000 tokens | $0.10 | | Cached input, up to 200,000 tokens | $0.004 | | Output | $0.50 | | Input, above 200,000 tokens | $0.30 | | Output, above 200,000 tokens | $1.50 | A request with 100,000 regular input tokens and 10,000 output tokens costs about **$0.015 — a cent and a half**. If all the input tokens qualify for the cached rate, the same request costs about **$0.0054 — about half a cent**. Compare that with the big models. Claude Fable 5.2, launched the same day, costs $8 per million input tokens and $40 per million output tokens — making Ember 80 times cheaper per token on both input and output. But the fairest comparison is with the model Ember is actually measured against: Anthropic's Claude Haiku 5.5. Ember's per-token prices are identical to Haiku 5.5's standard rate. The savings come from somewhere else. Independent benchmarking firm Artificial Analysis found that Ember matches Haiku 5.5's best score while using fewer tokens per job — so the same task costs about **$0.14 with Ember and $0.41 with Haiku 5.5**. Same price per token, one-third the bill. The model is cheaper because it is more efficient, not because the rate card is lower. There is a catch in the price table worth knowing. Above 200,000 input tokens, Ember's rates triple to $0.30 input and $1.50 output. For the long-document work Meta advertises, that jump matters. A million-token context window is only cheap for the first 200,000 tokens. AI agents have been expensive to run at scale, and that cost has quietly decided who gets to build them. An agent that makes hundreds of intermediate calls per task, at frontier-model prices, burns money faster than it creates value. That is why most "AI agent" products so far have been demos, or have been limited to a handful of large companies that can afford the bill. Ember attacks exactly that bottleneck. Meta's launch post positions it as a subagent that can handle hundreds of intermediate steps without sending every request to a larger frontier model. The pattern the industry is settling on is a division of labor: **the big model plans, the small model executes.** A frontier model decides what to do; a cheap model like Ember does the repetitive work of doing it — calling tools, parsing documents, running coding subtasks — hundreds of times per task, at a fraction of a cent each. When the cost of the repetitive layer falls this far, three things change at once: The demand side is already visible. [A Chinese AI agent app just hit 1.5 million daily users — and nobody in the West noticed](https://sheen6869.wixsite.com/shakeel-ahmed-ai-blo/post/a-chinese-ai-agent-app-just-hit-1-5-million-daily-users--and-nobody-in-the-west-noticed). People want agents that do things for them. Until now, the cost of running them has been the wall. Ember is a serious attempt to take that wall down. Cheap means nothing if the model cannot do the job. So the benchmark numbers matter, and Meta invited an independent check. Artificial Analysis, an independent evaluation firm, put Ember through its Intelligence Index, version 5, on the day of launch. At maximum reasoning effort, Ember scored **36 on the Artificial Analysis Intelligence Index**, effectively tied with Claude Haiku 5.5 at 36.2 (Ember itself measured 35.8) and well ahead of OpenAI's GPT-6 Luna, which scored 28. At a lower reasoning setting, Ember scored 34 at $0.12 per task. Artificial Analysis calls Ember the cheapest model on its leaderboard scoring 36 or higher. Meta's own claim is stronger: Ember "delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting," matching Haiku's best score at about a third of the cost and beating every Luna setting. Three honest caveats. First, these are benchmark estimates — your production bill will vary with workload shape, retry rate, and output length. Second, the Intelligence Index measures general capability; it does not prove Ember handles your specific task as well as Haiku does. Third, the gap to the top is real: Meta's own Muse Spark 1.3 scores 37, one point above Ember, but costs $2.39 per task — about 17 times more. Ember trades a little capability for a lot of savings, and that trade is the entire point. Ember is also the first model trained using Muse Wonder Full, Meta's newer training approach. Meta has expanded the Muse family fast since Muse Spark launched in April, followed by Muse Spark 1.3 and the Muse Wonder models in October. Ember is the bottom of that family by size — and, deliberately, the top of it by value. No model this cheap is a frontier model, and Ember does not pretend to be one. Its limits matter as much as its price. None of this is a reason to dismiss Ember. It is a reason to use it for what it is: the cheap, tireless worker inside a system where a bigger model still does the thinking. If you build software, the practical question is not whether Ember is the best model. It is whether your product can afford an agent layer now — and the answer changed on November 9. Start with the routing pattern. Use a frontier model for planning and judgment, and route the repetitive steps — tool calls, extraction, parsing, subtask execution — to Ember. That is the architecture Meta is selling, and it is the architecture that makes the economics work. [Claude can now read your bank account — and the agent wave behind it is the same one Ember is built for](https://aigrowthlabshakeel.blogspot.com/2026/11/claude-can-read-your-bank-account.html). If you are a solo developer, the more important number is $0.14 per task. That is the cost of letting an agent do a complete job — not a single call, a whole job. A hundred agent jobs a day costs $14. A thousand costs $140. Those are numbers a small business can budget for, and that is new. And if you are watching the job market rather than building products, the same shift cuts both ways. [AI layoffs in 2026 are hitting some roles while others are hiring — the map is more complicated than the headlines](https://dev.to/shakeelahmed/ai-layoffs-2026-which-jobs-are-actually-being-cut-and-which-ones-are-hiring-434g). Cheap agents do not replace the people who design the work. They replace the repetitive steps inside it — which is exactly the work Ember is built to do. The agent era has been waiting on one thing: a model cheap enough to run all day without asking permission. Ember is not the smartest model Meta has ever shipped. It is the cheapest one that is good enough. For most apps, that is the one that matters. **Sources:** [Meta — Introducing Our Smallest, Most Efficient Model: Muse Ember](https://research.meta.ai/blog/muse-ember) · [VentureBeat — Meta's Muse Ember nearly matches Claude Haiku 5.5 at one-third the estimated per-task cost](https://venturebeat.com/technology/metas-proprietary-muse-ember-nearly-matches-claude-haiku-5-5-at-1-3-the-estimated-per-task-cost) · [AlphaSignal — Meta Releases Muse Ember to Cut AI Agent Costs by 3x](https://alphasignal.ai/news/meta-releases-muse-ember-to-cut-ai-agent-costs-by-3x) **Read next on AI Growth Lab:** [A Chinese AI agent app just hit 1.5 million daily users — and nobody in the West noticed](https://sheen6869.wixsite.com/shakeel-ahmed-ai-blo/post/a-chinese-ai-agent-app-just-hit-1-5-million-daily-users--and-nobody-in-the-west-noticed) · [Claude can now read your bank account — should you let it?](https://aigrowthlabshakeel.blogspot.com/2026/11/claude-can-read-your-bank-account.html) · [AI layoffs 2026: which jobs are actually being cut — and which ones are hiring](https://dev.to/shakeelahmed/ai-layoffs-2026-which-jobs-are-actually-being-cut-and-which-ones-are-hiring-434g) *By Shakeel Ahmed — AI Growth Lab*