Muse Spark 1.1 benchmarks at 863 Elo on AA-Briefcase

• Meta-side posts put Muse Spark 1.1 at or near the top of clinical evals, with alexandr_wang's HealthBench post calling it SOTA on HealthBench Professional and _jasonwei's RadLE post saying it beats GPT-5.6 Sol and Gemini 3.1 on Radiology's Last Exam.
• Muse Spark 1.1 landed in the Debate Benchmark's top tier, where LechMazur's leaderboard post put it third overall behind Claude Fable 5 and Opus 4.7, ahead of GPT-5.6 Sol.
• AA-Briefcase showed a real agentic-work jump with a visible weakness: ArtificialAnlys's results thread reported 863 Elo, up 232 points over the previous Muse Spark, but 432 Presentation Elo, down 67 points.
• Coding-agent results looked competitive rather than leading, with ArtificialAnlys's coding-agent post placing Muse Spark 1.1 at 69 in the Opencode harness at roughly $1.4 per task.
• The cost story is per task, not just per token: ArtificialAnlys's cost thread says long agentic runs accumulate turns, tool calls, context, and cache effects.

Meta's launch post frames Muse Spark 1.1 as a multimodal reasoning model for agentic tasks, tool use, computer use, coding, and multimodal understanding. The developer docs list OpenAI and Anthropic SDK compatibility, parallel tool calls, streamed tool-call arguments, reasoning that carries across turns, and a 1M-token context window. Artificial Analysis's AA-Briefcase page says the benchmark spans 91 deliverable tasks across four multi-week knowledge-work projects, while Lech Mazur's Debate Benchmark repo uses side-swapped matchups, three-model judging, and Bradley-Terry ratings.

Meta paired Muse Spark 1.1 with the public preview of the Meta Model API, and the official product page lists $1.25 per 1M input tokens, $0.15 per 1M cached input tokens, and $4.25 per 1M output tokens.

The developer overview says the API exposes Responses, Chat Completions, and Messages formats under the same model, auth, and token pricing. The same docs list https://api.meta.ai/v1 as the base URL and muse-spark-1.1 as the model slug.

HealthBench Professional is an open benchmark for real clinician tasks, organized around care consult, writing and documentation, and medical research, according to the HealthBench Professional paper. Meta's posts claimed Muse Spark 1.1 is roughly at GPT-5.6 Sol's level on that benchmark, with _jasonwei saying it is similar or slightly better at lower cost.

Radiology's Last Exam 2.0 gave Muse Spark 1.1 two different-looking results: second behind Fable 5 on confidence-weighted score, first among AI models on handover readiness.

LechMazur described Debate Benchmark as adversarial, multi-turn debate across broad topics, where models need facts, rebuttals, coherence, and defensibility over several rounds. The benchmark repo says each matchup runs twice on the same topic with sides swapped, then a three-model judge panel decides winner and margin.

The model moves in LechMazur's leaderboard post were clean enough to scan:

• GPT-5.6 Sol: 1567 to 1684.
• Grok 4.5: 1410 to 1524.
• Sonnet 5: 1604 to 1622.
• MiniMax-M3: 1481 to 1541.
• Muse Spark 1.1: debuted at 1688, third overall.

Fable 5 still owned the board: LechMazur's side-swapped note said it won every side-swapped aggregate matchup, 130 out of 130.

AA-Briefcase is the benchmark behind the headline number. Artificial Analysis says it evaluates long-horizon knowledge work through deliverables such as spreadsheets, presentations, and memos, then combines rubric pass rate, analytical quality Elo, and presentation Elo.

Muse Spark 1.1's result split in ArtificialAnlys's thread was blunt:

• AA-Briefcase Elo: 863, up 232 points over the previous Muse Spark.
• Overall placement: roughly alongside Gemini 3.5 Flash and NVIDIA Nemotron 3 Ultra.
• Rubric pass rate: 34.5%, ahead of GPT-5.5 (xhigh) and just behind GLM 5.2 (max).
• Presentation Elo: 432, down 67 points from the previous Muse Spark.
• Lite subset behavior: deliverables were largely text-only, while top presentation models used more color, charts, and styling.

That is a very Meta-looking result: strong task execution, weak slide polish.

Artificial Analysis put Muse Spark 1.1 at 69 on its Coding Agent Index in the Opencode harness. The Coding Agent Index page says the score is a composite average pass@1 across DeepSWE, Terminal-Bench v2, and SWE-Atlas-QnA.

The comparison in ArtificialAnlys's coding-agent post was narrow but useful:

• Codex GPT-5.5 (medium): 71.
• Opencode Muse Spark 1.1 (xhigh): 69.
• Claude Code Opus 4.8 (medium): 67.
• Reported cost per task: about $1.4.
• Reported tradeoff: higher time per task.

Harness choice matters here. rasbt's harness caveat said GLM 5.2 and DeepSeek V4 were omitted from one chart because their available numbers came from Claude Code rather than a native harness, which would likely make them look worse.

Artificial Analysis argued that falling token prices do not automatically mean cheaper long-horizon agent runs. Complex tasks grow context, turns, tool calls, and reused inputs.

Their cost model in ArtificialAnlys's cost thread names four drivers:

• Token price: cheaper access can be offset by higher total token use.
• Turns: prior outputs become later inputs as the trajectory grows.
• Token efficiency: verbose models can cost more per task at lower list prices.
• Prompt caching and hit rate: reused instructions, research, tool outputs, and prior work change the economics.

Marketplace listing demo

alexandr_wang said Muse Spark can perform end-to-end tasks from short video instructions. rohanpaul_ai's demo showed the model extracting product details and photos from a mobile video, then operating a browser to fill and publish a Marketplace listing.

That demo lines up with Meta's launch framing around computer use, not just coding or chat.

OpenClaw v2026.7.1 added Muse Spark 1.1 alongside GPT-5.6 and broader UI, onboarding, mobile, and messaging-integration work.

Emdash v1.1.38 listed Muse Spark support with a new ACP-powered chat UI, Notion integration, Oh My Pi and Zero CLI support, storage management, draggable terminals in task tabs, macOS voice mode, and a task archive shortcut.