Model Drop

Muse Spark 1.2

August 5, 2026 2 min read

Muse Spark 1.2, which Meta released on August 5, 2026, is a coding-focused update to Muse Spark that shipped alongside Muse Code, Meta's new terminal coding agent.

Model: Muse Spark 1.2 (muse-spark-1.2)
Maker: Meta
Ship date: August 5, 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 beta on macOS and Linux
- Meta Model API, now beyond the US
When to reach for it:
- Budget coding agents in Muse Code
- High-volume tool calling on the cheap
- Skip it if you need open weights
How it compares:
- Trails Claude Opus 5 on coding
- Edges GPT-5.6 Terra and Grok 4.5 in places
What makes it different:
- Co-trained with its own coding harness
- Contributor tier trades data for price
Early reactions: Reviewers stressed that every launch number was Meta-run, mostly inside the harness the model was co-trained with, and independent tests showed smaller gains over 1.1.
Read more: Introducing Muse Code and Muse Spark 1.2 · Evaluation methodology

The benchmarks

Meta ran Muse Spark 1.2 and Muse Spark 1.1 at xhigh, Claude Opus 5 and GPT-5.6 Terra at max, and Grok 4.5 and Gemini 3.6 Flash at high; on Terminal-Bench 2.1 and DeepSWE v1.1 each model worked inside its own agent product (Muse Code, mini-swe-agent for 1.1, Claude Code, Codex, Grok Build, Antigravity). GDPval-AA v2 and MCP Atlas results come from Artificial Analysis and Scale AI, the coding bench is Meta's internal set, and the independent rows are today's scores.

Meta's numbers

Terminal-Bench 2.1 (agentic terminal tasks): 82.9%. −3.8 under Claude Opus 5 (86.7%), +1.1 over GPT-5.6 Terra (81.8%), +1.3 over Grok 4.5 (81.6%), +6.7 over Muse Spark 1.1 (76.2%).
DeepSWE v1.1 (long-horizon agentic coding): 59.3%. −5.7 under Claude Opus 5 (65.0%), −5.5 under GPT-5.6 Terra (64.8%), +2.7 over Grok 4.5 (56.6%), +19.3 over Gemini 3.6 Flash (40.0%).
Meta Internal Coding Bench (Meta's own codebase tasks): 70.6%. −8.8 under Claude Opus 5 (79.4%), +5.2 over GPT-5.6 Terra (65.4%), +2.3 over Muse Spark 1.1 (68.3%).
GDPval-AA v2 (office task Elo): 1631. −221 under Claude Opus 5 (1852), +54 over GPT-5.6 Terra (1577), +105 over Grok 4.5 (1526).
MCP Atlas (tool use over MCP): 90.3%. +4.5 over Claude Opus 5 (xhigh) (85.8%), +7.0 over Claude Fable 5 (83.3%), +8.5 over GPT-5.6 Sol (81.8%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2): 39.6. −11.2 under Claude Opus 5 (50.8), −2.5 under GPT-5.6 Terra (42.1), +0.8 over Grok 4.5 (38.8).
Terminal-Bench 2.1 (Artificial Analysis harness): 80.1%. −9.0 under Claude Opus 5 (89.1%), −7.9 under GPT-5.6 Terra (88.0%), −1.5 under Grok 4.5 (81.6%).
Epoch Capabilities Index (Epoch AI composite): 155. −8 under Claude Opus 5 (163), −5 under GPT-5.6 Terra (160), +1 over Grok 4.5 (154).