Table of Contents
Muse Ember: Meta's cheapest agent model, and when to actually run it
Three different labs just priced their cheap agent model at the exact same number: $0.10 per million input tokens, $0.50 per million output tokens. Anthropic got there first with Claude Haiku 5.5 on October 7, at that rate for prompts under 100,000 tokens per Anthropic. OpenAI matched it with GPT-6 Luna on September 22 per OpenAI. Meta landed on the identical rate card with Muse Ember, which shipped November 9 for prompts under 200,000 tokens per Meta AI Research. That is not a coincidence worth shrugging off. It means the sticker price stopped being the decision, and something else now is.
Muse Ember is Meta's smallest model, built for the part of agent work nobody designs around: the tool call, the routing step, the file read, the extraction pass that an agent runs hundreds of times in a session per Meta AI Research. Meta says it matches Claude Haiku 5.5's best benchmark score at roughly a third of the cost and beats every GPT-6 Luna reasoning setting on Artificial Analysis's cost-per-task measure. Those are Meta's numbers, not an independent audit, and that caveat matters more than usual here because all three labs are grading their own homework on adjacent benchmarks. What you can actually verify, cheaply, is the price match itself. What you can't yet verify is the ranking.
What it is
Ember rounds out a three-tier lineup: Muse Wonder for the hardest reasoning, Muse Spark 1.3 for coding and general agentic work, and now Ember underneath both for the repetitive stuff per Meta. It's the first model Meta trained using Muse Wonder Full as the teacher: the small model inherited judgment from the big one instead of being trained from scratch. It takes text, image, audio, and video input and returns text. Meta built it specifically for tool calls, coding subtasks, routing, extraction, and long-document parsing, not for the planning step that decides what an agent should do next.
Early access partners described narrow, repeatable jobs. Cline's founder called it "a fast, low-cost model they can rely on" for running as the subagent inside a coding harness per Meta's launch post. You.com reported a 2.2-second p95 latency on coding tasks, four to thirteen times faster than the open-weight models it compared against, with 36 to 48 percent fewer output tokens per task. Cloudflare's CTO said it ran "up to 30% faster at the tail" than the model it replaced in testing. Box used it to score contracts clause by clause against a negotiation playbook. None of those are controlled, third-party benchmarks, but they're a different kind of evidence than a vendor chart: specific partners, naming specific workflows, with numbers attached.
What it costs, and the number that actually matters
Here's the full rate card, straight from Meta's pricing page.
| Context size | Input / 1M tokens | Cached input / 1M tokens | Output / 1M tokens | Source |
|---|---|---|---|---|
| Up to 200K tokens | $0.10 | $0.004 | $0.50 | Meta AI Research |
| Over 200K tokens | $0.30 | $0.012 | $1.50 | dev.meta.ai pricing table |
Cross 200,000 tokens in a single call and the per-token price triples across the board, input, cached input, and output alike. If your agent is stuffing a full repo or a long PDF into context, which is exactly the use case Ember is built for, check which side of that line the call actually lands on before you trust a back-of-envelope cost estimate. A 1M-token context window makes it easy to forget you crossed the threshold.
Against Meta's own mid-tier model, the gap is bigger than the headline rate implies. Muse Spark 1.3 runs $1.25 input / $4.25 output per million tokens, roughly 12x Ember's input price. Meta's own chart puts the effective cost at $0.14 per Artificial Analysis task for Ember versus $2.39 for Spark 1.3, a 17x difference per task once you account for how many tokens each model burns getting to an answer per Meta. That per-task number, not the per-token rate, is the one worth tracking if you're actually running a volume workload.
If you'd rather not touch raw API billing, Meta also sells Ember through Muse Code subscriptions: Everyday at $5/month, High at $15, Power at $50 (which adds Muse Wonder Full access), and Maximum at $200, each tier buying a usage multiple over the one below it per dev.meta.ai. That's the flat-fee path if you don't want to manage a token budget at all.
How to turn it on
Two paths, depending on whether you want a CLI tool or raw API access.
For the API: Meta's Model API is OpenAI SDK-compatible, so you point your existing OpenAI client at dev.meta.ai and set the model string to muse-ember-1.0 per dev.meta.ai's quickstart docs. No new SDK to learn if you're already calling OpenAI-shaped endpoints.
For the CLI: install Muse Code with curl -fsSL https://dev.meta.ai/install.sh | bash on macOS or Windows, then log in through the Meta Model API portal per dev.meta.ai. Muse Code runs multiple agents in parallel by default, and Ember's natural slot is as the subagent model in that fan-out: Meta's own "agent fan-out" cookbook describes spinning up isolated worktrees for parallel subagent work, which is precisely the shape of job Ember is priced for.
It's also live on Google Cloud's Model Garden and Oracle Cloud, alongside Spark and Wonder, if you'd rather bill it through infrastructure you already pay for per Meta.
The limits worth knowing before you build on it
A 1M-token context window and up to 512K tokens of output is a large ceiling for a model this cheap, enough to take in a full codebase or document set in one call. But it's still an execution model, not a planner. Meta's own framing pairs Ember with Wonder or Spark for the hardest problems: Wonder or Spark plans, Ember handles the volume. Put it alone on a task that needs multi-step judgment and you're asking a subagent to do a lead's job.
On safety, Meta says it ran Ember through the Meta Superintelligence Scaling Framework, covering chemical and biological, cybersecurity, and loss-of-control risk categories, and that it stayed below the framework's high-risk thresholds across the categories assessed per Meta. That's a self-reported pass on an internal framework, not a third-party audit, which is worth remembering before you grant it tool access it doesn't need.
The benchmark claims, read carefully
Meta's headline numbers come from running each model at its best published harness, not the same harness for all three: mini-swe-agent for Ember, Claude Code for Haiku 5.5, Codex for Luna per Meta's footnotes. That's a meaningful asterisk. A harness can add or cost a model several points on its own.
| Benchmark | Muse Ember | Claude Haiku 5.5 | GPT-6 Luna | Source |
|---|---|---|---|---|
| Terminal-Bench 4.0 | 40.9% | 39.2% | 16.4% | Meta AI Research |
| AutomationBench-AA | 37.8% | 13.9% | 20.2% | Meta AI Research |
| Agents' Last Exam | 61.6% | 57.0% | 57.1% | Meta AI Research |
| Coding Agent Index-AA | 46 | 37 | 31 | Meta AI Research |
All scores at max reasoning effort, vendor-reported by Meta citing Artificial Analysis and Meta's internal evaluation suite, November 9, 2026. Not independently reproduced as of this writing.
Treat the ranking as Meta's claim, not a settled fact. The price match across three labs, by contrast, is easy to verify on each vendor's own pricing page. Artificial Analysis publishing Ember on its own leaderboard, under its own harness, would be the first real outside check on where it lands. Until then, the cost-per-task numbers are a vendor's chart, however specific.
The pick, by what stack you're already on
The three-way price match means lab loyalty, not cost, is now the real variable, so the pick comes down to where you already run agents and how many tokens a single call actually needs.
Already building on Claude Code or the Claude Platform: stay on Claude Haiku 5.5. It's the same $0.10/$0.50 rate up to 100,000 tokens, then $0.50/$2.50 above it per Anthropic, with no migration cost and Opus or Sonnet already in place to handle the planning Haiku skips.
Already on Codex or the OpenAI API: stay on GPT-6 Luna, live as gpt-6-luna in the API, ChatGPT, and Codex for paid tiers at the identical rate per OpenAI.
Starting fresh, already paying for Google Cloud or Oracle Cloud, or running calls that regularly brush past 100,000 tokens: run Muse Ember. Its cheap-tier window runs to 200,000 tokens, double Haiku's threshold, and Wonder or Spark sits right above it on the same infrastructure for the planning half of the job.
Hold the benchmark ranking loosely until Artificial Analysis publishes Ember on its own leaderboard, under its own harness, not Meta's. You can confirm the price match yourself in about three minutes by opening all three pricing pages side by side. The context-length threshold where each rate jumps, not the identical headline number every press release leads with, is what actually moves your bill at scale.