Muse Spark 1.2 is the coding-focused foundation model released by Meta Superintelligence Labs on August 5, 2026, co-trained with the Muse Code terminal agent it powers. It has a 1,048,576-token context window and accepts text, image, video and PDF input. On Meta's published charts it scores 82.9% on Terminal-Bench 2.1, 59.3% on DeepSWE v1.1 and 70.6% on Meta's internal coding benchmark, up from 76.2%, 53.0% and 68.3% for Muse Spark 1.1, placing it second behind Claude Opus 5 on all three. Pricing is $1.25 per million input tokens and $4.25 per million output tokens, with a data-contributor tier more than ten times cheaper. Meta publishes no parameter count or architecture details.
Parameters
Undisclosed
Context Window
1M
License
Proprietary
Release Date
2026-08-05
API Pricing
Input Price (per 1M tokens)
$1.25
Output Price (per 1M tokens)
$4.25
Billing Mode: per 1M tokens
Strengths
- •Near Opus-tier agentic coding at a fraction of the cost (82.9% on Terminal-Bench 2.1)
- •Jointly trained with the Muse Code harness for long-horizon, multi-turn workflows
- •1,048,576-token context with text, image, video and PDF input
Weaknesses
- •Places second behind Claude Opus 5 on all three of Meta's own coding benchmarks
- •Closed weights; no self-hosting or fine-tuning
- •No published parameter count or architecture details, limiting independent verification
Use Cases
- •Autonomous multi-step coding agents over large repositories
- •Multi-file change and refactoring automation
- •Long-running background engineering agents
Deep Analysis
Context Window
1,048,576 tokens
Accepts text, image, video and PDF input
API Price
$1.25 / $4.25 per 1M
Data-contributor tier ~10x cheaper ($0.10/$0.20)
Terminal-Bench 2.1
82.9%
With Muse Code; 2nd behind Claude Opus 5 (86.7%)
DeepSWE v1.1
59.3%
3rd, behind Opus 5 (65.0%) and GPT-5.6 Terra (64.8%)
Meta Internal Coding
70.6%
Up from 68.3% on Muse Spark 1.1
Weights
Proprietary
No self-hosting or fine-tuning; no param count published
Strengths
- ・Near Opus-tier agentic coding at a fraction of the cost (82.9% Terminal-Bench 2.1).
- ・Jointly trained with the Muse Code harness, so it shines on long-horizon, multi-turn workflows.
- ・1M context with text, image, video and PDF input for multimodal coding tasks.
- ・A data-contributor tier ~10x cheaper lowers the barrier for experimental use.
Weaknesses
- ・Places second behind Claude Opus 5 on all three of Meta own coding benchmarks.
- ・Closed weights: no self-hosting, fine-tuning, or independent architecture verification.
- ・No published parameter count, limiting external scrutiny.
Competitor Comparison
| Model | Arena | SWE | GPQA | Price |
|---|---|---|---|---|
| Muse Spark 1.2 | 82.9% (Terminal-Bench 2.1) | 59.3% (DeepSWE 1.1) | N/A | $1.25/$4.25 |
| Claude Opus 5 | 86.7% (Terminal-Bench 2.1) | 65.0% (DeepSWE 1.1) | N/A | $10/$50 |
| GPT-5.6 Terra | 81.8% (Terminal-Bench 2.1) | 64.8% (DeepSWE 1.1) | N/A | N/A |
| Grok 4.5 | 81.6% (Terminal-Bench 2.1) | N/A | N/A | N/A |
Muse Spark 1.2 is Meta Superintelligence Labs coding-focused foundation model, released August 5, 2026 and co-trained with the Muse Code terminal agent it powers. A 1M-token context with text, image, video and PDF input makes it a strong, comparatively cheap agentic-coding engine, second only to Claude Opus 5 on Meta own charts.
Sources
Analysis generated: 2026-09-19