Muse Spark 1.3

2026-09-02 · Publisher : Meta · Domain : Artificial Intelligence

Meta releases Muse Spark 1.3, its most advanced proprietary model, available in Muse Code and the Meta Model API, with gains in agentic tasks and reduced use of tools and tokens.

### Context Muse Spark 1.2 (August) had brought Meta closer to the frontier on code. Version 1.3 arrives a month later, while Meta has promised, but not yet published, open weights for 1.2. ### What the model brings - Better agentic and coding performance, according to Meta. - About 20% fewer tool calls and about 25% fewer tokens than version 1.2 (Meta claim). - Better collaborative work on long tasks and better awareness of its limits (Meta). - Increased robustness to adversarial attacks and prompt injection (Meta). ### Architecture and specifications Parameters, architecture, data and cutoff: not disclosed. Context: 1 M tokens and text, image and video input modalities according to Artificial Analysis and secondary sources. Availability: Muse Code and Meta Model API; a “max” variant in limited preview for partners, and the “xhigh” variant available. License: proprietary, weights not published. Identifier: “Muse Spark 1.3”. ### Cost and access | Tier | Input | Cached input | Output | |---|---|---|---| | Contributor | $0.10 | $0.002 | $0.20 | | Standard | $1.25 | $0.15 | $4.25 | Prices per million tokens, recorded from secondary sources (not read in the Meta post) and identical to those of version 1.2. Cost measured by Artificial Analysis: $0.55 per index task at xhigh, with peers of comparable score costing more than 70% more. ### Benchmarks | Intelligence Index (Artificial Analysis) | Score | |---|---| | Claude Fable 5.1 | 66 | | Claude Opus 5 (max) | 63 | | Muse Spark 1.3 (max, partner preview) | 62 | | Muse Spark 1.3 (xhigh) | 61 | | GPT-5.6 Sol (max) | 61 | | Grok 4.6 (high) | 61 | | Muse Spark 1.2 | 57 | | Muse Spark 1.1 | 53 | Source: Artificial Analysis (tiers), September 2026. On Tau3-Bench Banking, 47% at xhigh and 52% at max, versus 35% for the previous version. ### Versus closed models At the time of its release, the best closed models according to Artificial Analysis were Claude Fable 5.1 (66) and Claude Opus 5 at max (63). Muse Spark 1.3 (xhigh) ranks at 61, level with GPT-5.6 Sol (max) and Grok 4.6 (high), and at 62 at max; it gains 4 points over version 1.2. Meta itself compares it with GPT 5.6 Sol and Opus 5 in a table whose values have not been re-read. The measured cost per task is lower than that of peers of the same level, which is the main economic argument. On the open-versus-closed question, there are no open weights and no self-hosting, and the cheap tier implies use of the data by Meta. ### Key takeaways Muse Spark 1.3 places Meta among the leading models according to a third party, at a lower cost per task. The open/closed strategy remains hybrid: Glimmer is open, the top-of-range models are closed.

Tags : LLM, Meta, Muse, Poids fermés, Agents, Code, API