
The most important AI model release of the week costs about fourteen cents per task. Meta's Muse Ember, released on November 9, 2026, is not the smartest model Meta has ever shipped — it is deliberately the cheapest one that is still good enough. It matches Claude Haiku 5.5 on Artificial Analysis's Intelligence Index while costing roughly a third as much per task, and it exists for a single job: the repetitive work agents do all day. The news crossed into China's developer community within minutes — it was already sitting in LINUX DO's frontier-news feed, with developers weighing it against the relay-station economics they know intimately. The interesting question is not whether Ember beats Haiku. It is what happens to agent architecture when the execution layer gets this cheap.
Ember is the fourth model in Meta's Muse family, and the smallest. Meta's own framing gives it away: "Muse Wonder plans, Muse Ember handles the volume." Wonder, released in October, is the frontier reasoning model. Spark, from April, is the mid-weight coding model, with an open-source twin called Glimmer. Ember sits below all of them — not a downgraded flagship, but a model built from the start for high-volume, repeatable work: tool calls, coding subtasks, request routing, information extraction, and parsing long documents.
The numbers Meta published are striking precisely because they are boring. Standard pricing is $0.10 per million input tokens and $0.50 per million output tokens, with cached input at $0.004 per million. The context window is 1 million tokens, and it can generate up to 512,000 output tokens in a single call. It accepts text, images, audio, and video as input, and it is available through the Meta Model API and Muse Code under the model ID muse-ember-1.0. Meta does not disclose a parameter count, and the weights are not open — this is a hosted API product, not an open model release.
Meta also says Ember is its first model trained using Muse Wonder Full, its largest model, as the teacher. Distillation from a frontier teacher is how a small model punches above its weight on narrow tasks, and it explains the benchmark story: Ember scores 36 on the Artificial Analysis Intelligence Index at maximum reasoning, level with Claude Haiku 5.5, while Artificial Analysis estimates the cost per index task at $0.14 against Haiku's $0.41. Meta's own comparison adds that Ember beats OpenAI's GPT-6 Luna at every reasoning setting. Those are the vendor's claims, measured on a third-party benchmark suite — worth taking seriously, and worth remembering that benchmark tasks are not your production workload.
Ember on its own is a competent small model. Ember inside a routing architecture is a cost-structure change. The pattern is simple and increasingly standard in agent systems:
In this pattern, the planner might be 5% of the calls and the executor 95%. Cutting the executor's cost by two-thirds cuts the total bill by more than half, without touching quality on the steps that matter. That is why Meta's marketing talks about "cost per task" instead of "cost per token" — the unit that matters to an agent operator is the completed job, not the token.
Early testers named by Meta include Cloudflare and You.com, which is a telling detail: both run agents at internet scale, where the difference between $0.41 and $0.14 per task is the difference between a feature that ships and a feature that stays in a slide deck.
If you want to try the model, the entry point is the Meta Model API:
curl https://api.meta.ai/v1/chat/completions \
-H "Authorization: Bearer $META_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "muse-ember-1.0",
"messages": [{"role": "user", "content": "Extract the invoice totals from this document."}]
}'
The model ID and pricing above are from Meta's announcement as reported by TechCrunch and other outlets on November 9. As with any new API, check the current docs before wiring it into production — pricing and availability for a model this new can move.
The LINUX DO thread that surfaced Ember in China had no discussion yet when checked — it was minutes old. But the community's reaction pattern is easy to predict, because this forum has spent years dissecting the economics of cheap AI access. The same readers who recently picked apart the relay-station business — third-party services reselling API access below official prices — will read Ember's $0.10-per-million-input price as a direct answer to a question they have been asking for years: when does the official price get low enough that the gray market stops making sense?
Ember does not kill relay stations by itself. But it is part of a visible trend: frontier labs are now competing on the price of the execution tier, not just the capability of the flagship. Anthropic cut Fable 5.2's API price by 20% the same day Ember launched. When the official small model costs $0.14 a task, the margin for reselling someone else's quota gets very thin — and the risk of bans, throttling, and sudden policy changes, which the relay-station postmortem documented in detail, stops being worth it.
There is a second reason the price matters in China specifically. Developers there have long routed around dollar-denominated API costs, foreign credit cards, and regional access limits. A model that is cheap enough to run at scale, through an official API, changes the calculation for indie hackers building agent products: the infrastructure question moves from "how do I afford the tokens" to "how do I design the routing."
Small models are not a universal replacement, and Ember's own spec sheet says so. A 1-million-token context window sounds enormous until you fill it with a whole codebase and a long conversation history; retrieval quality degrades long before the window fills. And a model trained for repetition will still fail on tasks that need genuine reasoning — that is what the planner tier is for. The honest way to use Ember is as a specialist, not a generalist:
| Workload | Use Ember? | Why |
|---|---|---|
| Tool calls and API routing | Yes | Repetitive, structured, easy to verify |
| Extracting data from documents | Yes | Long-context parsing is the stated strength |
| Coding subtasks and test runs | Yes | Bounded scope, fast feedback loop |
| Open-ended planning and design | No | Needs the frontier model's judgment |
| Novel reasoning and research | No | This is the Wonder/Fable tier |
The failure mode to watch for is silent degradation: a small model that confidently returns a slightly wrong extraction, or a subtly broken test, repeated hundreds of times a day. The routing pattern only works if the planner tier actually verifies outcomes. Teams that adopt Ember without a verification step will save money and lose quality — the exact tradeoff the plan-mode pattern exists to prevent.
Step back, and Ember is less a product launch than a signpost. The agent era has a cost problem: an agent that makes a hundred model calls to finish one job is a hundred times more expensive than a chatbot that makes one. The industry's answer, arriving from every lab at once, is tiered intelligence — a frontier model that thinks, and a fleet of cheap models that do. Meta's lineup now maps cleanly onto that idea: Wonder plans, Spark codes, Ember executes.
For English-speaking developers, the practical takeaway is concrete: if you are building agents on a single model today, you are almost certainly overpaying for the repetitive 95% of the work. The routing pattern is no longer exotic infrastructure — it is the default architecture the vendors are designing for, and the price gap between tiers is now large enough to justify the engineering. Start by identifying the steps in your agent's loop that are mechanical and verifiable, and route those to the cheapest competent model you can find. Ember, at $0.14 a task, is currently the cheapest credible option on that list. It will not be the cheapest for long, which is itself the point: the execution layer is becoming a commodity, and the value is moving up the stack to whoever designs the best routing.
Check the price before you build on it, verify the benchmark claims against your own workload, and treat the model as what Meta says it is — the volume worker, not the brain. Used that way, it is the most useful kind of release: not a new capability, but a new price for an old one.