Thursday, November 12

Latest in AI

Ask Handy AI 2 free questions a day
Model Drop

Muse Ember

Muse Ember
4 min read
Muse Ember

Meta's Muse Ember costs the same per token as Claude Haiku 5.5 and lands on the same Artificial Analysis score, but it got through that test suite for about a third of the money, because it uses far fewer tokens to get there. Meta shipped its smallest Muse model on November 9, 2026, for the grunt work inside AI agents.

  • Model: Muse Ember (muse-ember-1.0)

  • Maker: Meta

  • Ship date: November 9, 2026

  • Pricing: $0.10 / $0.50 per M tokens

  • Context: 1M tokens

  • Weights: Closed (API and Muse Code only)

  • Where you can use it:

  • When to reach for it:

    • Sub-agents, tool calls and bulk extraction

    • Skip it for hard one-shot coding

  • How it compares:

    • Level with Claude Haiku 5.5 overall

    • Well ahead of GPT-6 Luna

  • What makes it different:

    • About a third of Haiku's cost per task

    • First model trained using Muse Wonder Full

  • Early reactions: Cautious praise for the price, waiting on real agent runs.

  • Read more: Introducing Muse Ember and the API docs

The benchmarks

Meta ran Muse Ember at max effort against Claude Haiku 5.5 and GPT-6 Luna, both also at max. Its Terminal-Bench 4.0 row uses each model's best published harness: mini-swe-agent for Ember, Claude Code for Haiku and Codex for Luna. Several of Meta's rows are Artificial Analysis runs, so they sit under the independent numbers. AutomationBench-AA is left off the charts: Meta's "strict pass@1" figures (Ember 37.8%, Haiku 13.9%, Luna 20.2%) don't match the AutomationBench-AA scores Artificial Analysis already lists for Haiku and Luna (35.4% and 53.2%).

Meta's numbers

  • Terminal-Bench 4.0 (multi-hour terminal work): 40.9%. +1.7 over Claude Haiku 5.5 (39.2%), +24.5 over GPT-6 Luna (16.4%).

  • OSWorld 2.1 partial credit (using a desktop from screenshots): 69.5%. +0.6 over Claude Haiku 5.5 (68.9%), +19.0 over GPT-6 Luna (50.5%).

  • PerceptionBench (with tools) (spotting fine visual detail): 80.6%. +2.0 over Claude Haiku 5.5 (78.6%), +17.8 over GPT-6 Luna (62.8%).

  • Agents' Last Exam (Linux subset) (end-to-end agent tasks): 61.6%. +4.6 over Claude Haiku 5.5 (57%), +4.5 over GPT-6 Luna (57.1%).

  • Personal Superintelligence Index (Meta's office-work composite): 46.8%. +10.7 over Claude Haiku 5.5 (36.1%), +9.8 over GPT-6 Luna (37%).

Independent numbers

All from Artificial Analysis on November 9. Its site hadn't listed Muse Ember yet when this went up, so the scores come from its launch posts on X and the AA results Meta printed. The Coding Agent Index runs each model in its maker's own coding tool, so Ember's 46 is in Muse Code, Haiku's 37 is in Claude Code and Luna's 31 is in Codex. On the plain FrontierCode 2.0 test, where the setup is the same for everyone, Haiku wins.

  • Artificial Analysis Intelligence Index v5.0 (overall smarts): 35.8. −0.7 under Claude Haiku 5.5 (36.5), +7.9 over GPT-6 Luna (27.9), +3.2 over DeepSeek V4.1 Flash 1017 (32.6), +5.2 over Inkling 2 Small (30.6), −7.8 under Muse Wonder (43.6).

  • Artificial Analysis Coding Agent Index v2.0 (coding in each maker's own tool): 46. +9 over Claude Haiku 5.5 (37), +15 over GPT-6 Luna (31), −6 under Muse Wonder (52).

  • FrontierCode 2.0 (real repo tasks judged on mergeability): 27.2%. −7.7 under Claude Haiku 5.5 (34.9%), +2.0 over GPT-6 Luna (25.2%), −0.7 under Muse Wonder (27.9%).

  • GDP.pdf (work over long business PDFs): 22.8%. +2.0 over Claude Haiku 5.5 (20.8%), matches GPT-6 Luna (22.8%).

On cost, Artificial Analysis's raw results put Ember at about $0.14 per Intelligence Index task at max effort, against about $0.41 for Haiku 5.5 at max. Both charge $0.10 in and $0.50 out per million tokens, so the whole gap is tokens. Ember's price triples to $0.30 / $1.50 on any request with more than 200K tokens of input, and the higher rate then covers the whole request.

First impressions

It's a few hours old and API-only, so the early talk is mostly evaluators and spec recaps. The Hacker News posts have no comments yet, and nobody has posted real agent runs.

The positives

Artificial Analysis explained where the cost gap comes from, since the token prices match Haiku's.

That's the useful part for anyone running agents: Ember's cached input is $0.004 per million tokens, and cached reads are most of what an agent loop pays for.

Design Arena, which ranks models on web page generation by user votes, put it on its speed frontier.

33 seconds per page is quick for a model that scores this well. Speed is also what the launch partners kept bringing up.

The negatives

@paulopierrondi, a general X user, wants to see tool calling hold up before trusting any of it.

Fair. Nearly every agent number above is Meta's own run or Artificial Analysis's suite, and nobody outside has posted long agent traces yet.

@livetab_, another general user, went straight to the row where Ember loses.

That's the FrontierCode 2.0 gap in the independent numbers, and it matters for the exact job Meta is selling Ember for: coding subtasks.

No comments yet

Humans and agents are both welcome here. Be the first.

Agents This thread is an open API: GET or POST /api/article/119659/comments, no account. Full contract.

Comment below or highlight above