Model Drop

Muse Spark 1.3

September 2, 2026 2 min read

Muse Spark 1.3, which Meta released on September 2, 2026, is an agentic and coding upgrade to Muse Spark that shipped at xhigh reasoning, with its max-reasoning mode held in limited preview for extra safety testing.

Model: Muse Spark 1.3 (muse-spark-1.3)
Maker: Meta
Ship date: September 2, 2026
Pricing: $1.25 / $4.25 per M tokens ($0.10 / $0.20 on the contributor tier, where Meta may train on your data)
Context: 1M tokens
Where you can use it:
- Muse Code on macOS and Linux
- Meta Model API
When to reach for it:
- Long, multi-step agent jobs
- Very long documents and codebases
- Skip it if you need open weights
How it compares:
- Behind Claude Fable 5.1 and Claude Opus 5
- Level with GPT-5.6 Sol on coding
What makes it different:
- Asks before unclear or risky steps
- Fewer tool calls and tokens than 1.2
Early reactions: Early testers liked the output and speed but called Muse Code rough, and reviewers warned that Meta's headline max scores weren't what most users could run yet.
Read more: Introducing Muse Spark 1.3 · Evaluation methodology

The benchmarks

Meta's table runs Muse Spark 1.3, Claude Opus 5 and GPT-5.6 Sol at max and Muse Spark 1.2 at xhigh, taking each rival's best score from Meta's runs, official leaderboards or self-reports. Max was held back for extra safety testing at launch, so the Artificial Analysis rows use the public xhigh model, and the independent rows are today's scores.

Meta's numbers

GDPval-AA v2 (office task Elo): 1754. −70 under Claude Opus 5 (1824), +44 over GPT-5.6 Sol (1710), +139 over Muse Spark 1.2 (1615).
JobBench (professional tool use): 64.9. −0.8 under Claude Opus 5 (65.7), +19.5 over GPT-5.6 Sol (45.4).
OSWorld 2.0 (computer use, partial credit): 66.9%. −1.4 under Claude Opus 5 (68.3%), +4.2 over GPT-5.6 Sol (62.7%).
DeepSearchQA (agentic web search): 90.3%. −2.8 under GPT-5.6 Sol (93.1%), −0.1 under Claude Opus 5 (90.4%).
AutomationBench (end-to-end business workflows): 49.6%. −0.7 under Claude Opus 5 (50.3%), +2.9 over GPT-5.6 Sol (46.7%).
MRCR 512K-1M (million-token recall): 98.1%. +24.3 over GPT-5.6 Sol (73.8%), +42.6 over Muse Spark 1.2 (55.5%).
DeepSWE v1.1 (long-horizon agentic coding): 75.4%. +1.4 over Claude Opus 5 (74.0%), +2.4 over GPT-5.6 Sol (73.0%), +20.4 over Muse Spark 1.2 (55.0%).
Terminal-Bench 2.1 (agentic terminal tasks): 88.8%. +0.0 over GPT-5.6 Sol (88.8%), +2.1 over Claude Opus 5 (86.7%), +5.9 over Muse Spark 1.2 (82.9%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2, xhigh): 45.1. −5.7 under Claude Opus 5 (50.8), −1.9 under GPT-5.6 Sol (47.0), +4.2 over Gemini 3.8 Flash (high) (40.9).
Terminal-Bench 2.1 (Artificial Analysis harness, xhigh): 85.4%. −3.7 under Claude Opus 5 (89.1%), −2.6 under GPT-5.6 Sol (88.0%), −2.2 under Gemini 3.8 Flash (high) (87.6%).
Epoch Capabilities Index (Epoch AI composite): 157. −6 under Claude Opus 5 (163), −5 under GPT-5.6 Sol (162), +0 over Gemini 3.8 Flash (157).