Model Drop

Muse Spark

April 8, 2026 2 min read

Muse Spark, which Meta launched on April 8, 2026, is the first model from Meta Superintelligence Labs, a closed multimodal reasoning model built to power the Meta AI app.

Model: Muse Spark (in the Meta AI app and meta.ai; no public API)
Maker: Meta
Ship date: April 8, 2026
Pricing: Free in Meta AI (API in private preview only)
Context: 262K tokens (per Artificial Analysis; Meta didn't publish one)
Where you can use it:
- meta.ai and the Meta AI app
- Private API preview for select partners
When to reach for it:
- Health questions and reading charts
- Photo help, like fixing an appliance
- Skip it for coding and long agent runs
How it compares:
- Behind Gemini 3.1 Pro and GPT-5.4 overall
- Close to Claude Opus 4.6, ahead of Grok 4.20
What makes it different:
- Contemplating mode runs parallel agents
- Health data curated by 1,000+ physicians
Early reactions: Reviewers called it genuinely competitive but noted Gemini 3.1 Pro still led most categories, and Meta itself admitted gaps in coding and long agent work.
Read more: Introducing Muse Spark · Safety & Preparedness Report

The benchmarks

Meta tested Muse Spark in Thinking mode, reporting whichever was higher of each rival's self-reported score or Meta's own API rerun; ARC-AGI-2 is the public set at pass@2, and the separate Contemplating mode, still rolling out, isn't in these rows. Rivals ran at their top settings: Claude Opus 4.6 at max, Gemini 3.1 Pro at high, GPT-5.4 at xhigh and Grok 4.20 in reasoning mode. The independent rows are today's scores.

Meta's numbers

HealthBench Hard (open-ended health queries): 42.8%. +2.7 over GPT-5.4 (40.1%), +22.2 over Gemini 3.1 Pro (20.6%), +28.0 over Claude Opus 4.6 (14.8%).
CharXiv Reasoning (scientific figure questions): 86.4%. +3.6 over GPT-5.4 (82.8%), +6.2 over Gemini 3.1 Pro (80.2%), +21.1 over Claude Opus 4.6 (65.3%).
DeepSearchQA (agentic web search): 74.8%. +1.1 over Claude Opus 4.6 (73.7%), +1.2 over GPT-5.4 (73.6%), +5.1 over Gemini 3.1 Pro (69.7%).
Humanity's Last Exam (no tools): 42.8%. −2.6 under Gemini 3.1 Pro (45.4%), −1.1 under GPT-5.4 (43.9%), +2.8 over Claude Opus 4.6 (40.0%).
GPQA Diamond (PhD-level science questions): 89.5%. −4.8 under Gemini 3.1 Pro (94.3%), −3.3 under GPT-5.4 (92.8%), −3.2 under Claude Opus 4.6 (92.7%).
SWE-bench Verified (real GitHub fixes): 77.4%. −3.4 under Claude Opus 4.6 (80.8%), −3.2 under Gemini 3.1 Pro (80.6%).
Terminal-Bench 2.0 (agentic terminal tasks): 59.0%. −16.1 under GPT-5.4 (75.1%), −9.5 under Gemini 3.1 Pro (68.5%), −6.4 under Claude Opus 4.6 (65.4%).
ARC-AGI-2 (abstract reasoning puzzles): 42.5%. −34.0 under Gemini 3.1 Pro (76.5%), −33.6 under GPT-5.4 (76.1%), −20.8 under Claude Opus 4.6 (63.3%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2, AA estimate): 31.3. −0.6 under Claude Opus 4.6 (31.9), +1.6 over Gemini 3.1 Pro (29.7), −7.7 under GPT-5.4 (39.0).
Epoch Capabilities Index (Epoch AI composite): 152. −3 under Claude Opus 4.6 (155), −3 under Gemini 3.1 Pro (155), −5 under GPT-5.4 (157).