Muse Ember for Voice and Chat Agents: Tool Calls, Latency, and How to Test It

Meta shipped Muse Ember on November 9, 2026, and priced it at $0.10 per million input tokens. It is Meta's smallest Muse model, built for the unglamorous work your agent does on repeat, like calling tools, routing requests, pulling a record, parsing a long document.

That happens to be the layer where most production voice and chat agents spend their tokens, so Ember is worth a hard look. The catch is that the numbers Meta published are agentic benchmarks. None of them score tool-call accuracy or voice.

Its refusal behavior helps in one scenario and hurts in another, and the only speed figures so far come from hand-picked launch partners.

Here is what the launch data says about agent work, where it goes quiet, and how to test Ember on your own calls before you move any traffic.

Muse Ember at a Glance

• Released: November 9, 2026. Model ID muse-ember-1.0.

• Built for: tool calls, coding subtasks, routing, extraction, and long-document parsing, per Meta.

• Price: $0.10 input, $0.004 cached input, $0.50 output, per million tokens, for prompts up to 200K tokens.

• Context: 1M tokens in, up to 512K tokens out.

• Inputs: Meta calls Ember natively multimodal (text, image, audio, and video). Meta’s Model API docs list text, image, video, audio, and PDF in, text out, but Artificial Analysis lists only text and image. Test audio input on your own calls before you build a voice flow on it.

• API: the Meta Model API, with Responses, Chat Completions, and Messages endpoints. Also wired into Meta’s Muse Code agent.

• Intelligence: 36 on the Artificial Analysis Intelligence Index at max reasoning, tied with Claude Haiku 5.5, for roughly a third of the cost per task.

Why a Small Model Matters for Agents

Most agent turns don't need a frontier model. Confirming an appointment, looking up an order, filling one form field, or deciding which department a caller wants is short, repeatable work.

What you actually need there is the right tool, the right arguments, a fast answer, and a bill that stays small when call volume climbs. That’s the job Meta built Ember for.

Meta frames it as the executor in a two-tier setup. In the words of Meta's launch post, you pair it with Muse Wonder: "Muse Wonder plans, Muse Ember handles the volume." Wonder decides what to do, and Ember runs the many small steps.

For agent teams, that turns model choice into a per-step decision. And every planner-to-executor handoff is one more place a call can quietly break.

What the Benchmarks Say About Agent Work

Meta's table puts Ember against Claude Haiku 5.5 and GPT-6 Luna, all at max reasoning. Several rows are Meta's own internal evaluations, so read them with that in mind.

Read together, this is as a mixed result. Ember’s clearest wins are multistep workflows and coding in real agent harnesses.

On Terminal-Bench and the OSWorld partial score, it’s within about two points of Haiku. Haiku 5.5 takes FrontierCode by 7.7 points, Luna ties Ember on long documents, and the overall index is a dead heat.

Meta’s chart labels its AutomationBench figure strict pass@1, while Artificial Analysis’s own write-up puts Ember at 68.1% on the same benchmark.

Terminal-Bench ran each model on its own best harness: mini-swe-agent for Ember, Claude Code for Haiku, and Codex for Luna.

None of these, though, is a voice or tool-calling benchmark. They don't check whether Ember passes a correctly formatted date to your booking API, waits for the tool result before it answers, or keeps the caller company while a lookup runs.

Cekura’s own benchmark of voice agent platforms shows how far agent performance can swing from one platform to the next.

Ember Says "I Don't Know" a Lot (And That Cuts Both Ways)

On Artificial Analysis's AA-Omniscience factual-recall test, Ember declined to answer 64% of questions. It answered 21.1% correctly, below Haiku 5.5 at 34.7%, and its hallucination rate of 17.4% came in at about half of Haiku's.

For a voice or chat agent, picture both sides of that on a live call.

The upside is safety. An agent that won't invent a refund policy, a price, or an open appointment slot is the agent you want when the facts are supposed to come from a tool or your knowledge base.

The downside shows up when your prompt or retrieval leaves a hole. A model this quick to decline will stall, deflect, or hand the call off. On the floor, that reads as longer calls, more transfers, and lower containment, and none of it trips an error log.

So test both halves. Run scenarios where the answer lives in the tool result, and scenarios where it doesn't, then check that the agent escalates instead of looping.

Meta also says it evaluated Ember for seeking human input before high-impact actions, which is worth holding against your own transfer and confirmation rules.

Latency: What's Been Claimed So Far

Meta hasn't published a time-to-first-token figure for Ember. The speed numbers so far come from launch partners Meta chose to quote, so treat them as vendor-selected:

• Cloudflare: up to 30% faster at the tail than the comparable model it tested, on multi-repo investigation tasks.

• You.com: a p95 of 2.2 seconds on its coding tasks, 4x to 13x faster than leading open-weight models, and 36% to 48% fewer output tokens.

None of those ran in a voice pipeline. A 2.2-second p95 is a long silence on a phone call if it lands in front of every turn.

What you need is the time to first token, and then first audio, for your prompt size, your reasoning setting, and your region, measured inside your full stack. Our voice AI latency guide walks through how to budget each step.

Reasoning setting changes the math too. Ember's headline scores are at max. Meta's own cost chart puts a lower setting near 34 on the Artificial Analysis index for about $0.12 per task. Lower settings should answer sooner, but they can call tools less reliably, so test the exact setting you plan to ship.

Cost at Agent Scale

Ember, Haiku 5.5, and GPT-6 Luna all list the same headline rate of $0.10 input and $0.50 output per million tokens.

The difference is in tokens per task. Artificial Analysis measured about $0.14 per Intelligence Index task for Ember at max, against roughly $0.41 for Haiku 5.5 at the same score.

These two details move the bill most for agents:

• Cached input runs $0.004 per million tokens, roughly a 96% discount. Agents resend the same system prompt, tool definitions, and policy text on every single turn, so that discount lands squarely on your highest recurring cost.

• Fewer output tokens per turn. If You.com's 36% to 48% figure holds for your traffic, that trims both spend and the time needed to generate each reply.

It’s worth knowing these rates cover prompts up to 200K tokens. Past that, Meta lists Ember at $0.30 input, $0.012 cached input, and $1.50 output per million tokens, and Artificial Analysis notes Haiku’s rate climbs on very long prompts too.

Price high-context workloads at the long-prompt rate.

Tool-Calling Migration Checklist

If you're thinking about swapping your current model for muse-ember-1.0, run these checks before any live traffic moves:

1. API format. Ember supports the Responses, Chat Completions, and Messages formats, and Meta’s docs list tool calling and structured outputs. Confirm your framework’s tool-call schema and streaming events map cleanly, and verify parallel tool calls and forced tool choice in Meta’s API docs.

2. Tool argument formats. Dates, phone numbers, currency, and IDs are where small models slip. Validate the arguments against your schemas, not just that a tool got called.

3. Reasoning setting. Ember defaults to max. Pick the setting you plan to ship, then measure latency there.

4. Refusals and escalations. Given the high decline rate, measure how often Ember transfers or deflects compared with your current model.

5. Planner-to-executor handoffs. If you pair Ember with a larger planner, test the handoff both directions: does Ember get enough context to finish the step, and does the planner catch what Ember misses?

How to Test Muse Ember Before You Switch Your Agent

A cheaper model only saves you money if it finishes the same calls. The benchmarks say nothing about your tools, your prompts, or your callers, so the only test that counts is the one you run on your own traffic.

That is the work Cekura is built for. It runs end-to-end conversational simulations and scores them, so you can compare your current model and Ember on the calls that pay your bills, before you cut over.

Our guide to testing new model versions on real production calls covers the method. For Ember, focus the comparison on four things:

1. Baseline first. Run your current model on a fixed scenario set and record task completion, tool-call accuracy, transfers, and response time.

2. Tool-call correctness. Score each call for the right tool, valid arguments, and correct use of the result. An LLM as a judge can grade this at scale, with humans spot-checking the edges.

3. Knowledge gaps. Include scenarios where the answer isn't available, and confirm the agent escalates instead of looping or inventing one.

4. Latency under load. Measure time to first audio at realistic concurrency, then track containment, transfers, and drop-offs after rollout. Our voice AI evaluation metrics guide lists what to watch.

With Cekura, teams can:

• Simulate before launch: run thousands of caller scenarios against your current model and Ember, then compare tool-call accuracy and task completion side by side.

• Stress the infrastructure: replay interruptions, background noise, and poor audio, where slow responses and missed turns start to bite.

• Monitor production calls: track drop-off points, sentiment, workflow adherence, and compliance failures on live traffic.

• Replay known failures: re-run the exact call that broke against your updated agent to confirm the fix.

Native integrations work out of the box for Retell, VAPI, ElevenLabs, LiveKit, Pipecat, Bland, and more. You don't rebuild anything. You add a testing and monitoring layer on top of what you already have.

Cekura supports SOC 2, HIPAA, and GDPR compliance, covering transcript redaction, role-based access, and audit trails. Healthcare teams run these tests under HIPAA, where a missed branch can mean a patient misses an appointment.

Book a demo to test Muse Ember on your own calls.

Frequently Asked Questions

What is Muse Ember?

Muse Ember is Meta's smallest and most efficient Muse model, released November 9, 2026. Meta built it for high-volume agent work like tool calls, routing, extraction, and coding subtasks. It's available through the Meta Model API as muse-ember-1.0 and inside Muse Code.

Is Muse Ember good for tool calling?

Muse Ember leads Claude Haiku 5.5 and GPT-6 Luna on most agentic benchmarks in Meta's launch table, including AutomationBench-AA and Terminal-Bench 4.0.

None of those tests measure tool-call accuracy inside a live voice or chat agent, so test it on your own tools and scenarios before you rely on it.

Is Muse Ember fast enough for voice agents?

Meta hasn't published a time-to-first-token figure for Muse Ember. Launch partner You.com reported a p95 latency of 2.2 seconds, and Cloudflare reported up to 30% faster tail latency, both on coding work, outside a voice pipeline.

Measure it inside your own stack at the reasoning setting you plan to run.

How much does Muse Ember cost?

Muse Ember costs $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens for prompts up to 200K tokens, with higher rates above that.

Artificial Analysis measured about $0.14 per task on its Intelligence Index at max reasoning, roughly a third of Claude Haiku 5.5's $0.41 at the same score.

Does Muse Ember hallucinate less?

Artificial Analysis measured a 17.4% hallucination rate for Muse Ember, about half of Claude Haiku 5.5's. Ember also declined to answer 64% of factual-recall questions, so test whether your agent escalates correctly when it doesn't have an answer.

Muse Ember vs Claude Haiku 5.5: which is better for agents?

Neither one wins everything. Muse Ember ties Haiku 5.5 at 36 on the Artificial Analysis Intelligence Index for about a third of the cost per task and leads most agentic tests in Meta's table, while Haiku 5.5 is ahead on FrontierCode 2.0 coding and factual recall.

Run both on your own scenarios before you choose.