Meta Unveils Muse Spark 1.3 With Major Coding Boost and Parity With GPT-5.6 Sol at Low Prices
·
03.09.2026 22:12
·
3 min read
Meta has released a major update to its flagship AI model, Muse Spark 1.3. The update focuses on improving coding capabilities, executing agentic workflows, and maintaining context over extended sessions. The model is already available to developers through the Muse Code environment and the Meta Model API.
In a blog post on Meta AI Research, the company notes that in software development tasks, the algorithm uses an average of 25% fewer tokens and makes 20% fewer intermediate third-party tool calls compared to version 1.2, generating more concise and accurate code.
Benchmark Results
According to independent testing by analysis platform Artificial Analysis, the publicly available Muse Spark 1.3 xhigh configuration scored 61 points on the composite Intelligence Index, matching the top-tier GPT-5.6 Sol.
The non-public Muse Spark 1.3 max variant with an expanded reasoning module scored 62 points, reaching the level of Claude Fable 5. However, the model still falls short of Anthropic’s latest Fable 5.1 (65–66 points).
In specialized coding benchmarks, Meta’s model demonstrated a clear lead: in the DeepSWE 1.1 benchmark, version 1.3 max scored 75.4% compared to 72.7% for GPT-5.6 Sol. As VentureBeat points out, the public release of the maximum reasoning mode is delayed while final safety checks are completed.
Pricing and Open Weights
The main competitive advantage of the lineup remains its query pricing. According to Artificial Analysis, the standard rate is set at $1.25 per million input tokens and $4.25 per million output tokens (cache reads cost $0.15). The context window spans 1 million tokens and supports text, image, and video processing.
Developers can also opt for an ultra-low-cost Contributor tier ($0.10 input and $0.20 output), though by selecting it, Meta retains the right to use conversation data to train future systems.
Muse Spark 1.3 is rolling out today with frontier performance almost too cheap to meter. This is the biggest jump we've made so far on coding and agentic work. Try it in Muse Code and our API.
— Mark Zuckerberg (@finkd) September 2, 2026
Next up 🍉 and Muse Spark open weights releases coming soon. pic.twitter.com/XQQEDEJGD7
The company confirmed plans to eventually release an open-weights version of the model, though it has not yet disclosed which specific iteration will be made publicly available.
