On September 2, 2026, Meta Platforms officially released its most powerful AI model to date, Muse Spark 1.3. The model is designed to extend agent workflows and enhance coding capabilities, and is now available in Muse Code and the Meta Model API. Meta Chief AI Officer Alexandr Wang said Muse Spark 1.3 surpasses OpenAI's GPT-5.6 Sol in coding ability and performs on par with Anthropic's Claude Fable 5.1, reaching state-of-the-art overall.

Details: Capability Upgrades and Efficiency Optimizations

Muse Spark 1.3 is the latest iteration of the Muse series (the 1.2 version was released about a month earlier), focusing on two main areas:

Long-horizon agent workflows

The model is designed to collaborate with users within a single long thread, handling multiple workflows in parallel. When faced with open-ended goals, it can proactively use tools to construct context from messy or conflicting information sources, repair planning gaps, and continuously track learned content to deliver complete results. This capability directly serves Meta's long-term vision for "personal agents" that can act on a user's behalf 24/7.

Coding capabilities

Additional training was applied for longer-horizon coding tasks to improve usability in common engineering workflows. Compared with version 1.2, the model is more restrained in unnecessary turns, produces more concise outputs, and yields cleaner code style. Meta engineers report about a 20% reduction in tool calls, roughly 25% lower token consumption, and faster, more efficient runs.

On public benchmarks, Muse Spark 1.3 (max variant) performed strongly across coding and agent tests: DeepSWE v1.1 scored 75.4 (surpassing GPT-5.6 Sol's 73 and Opus 5's 74), Terminal-Bench 2.1 reached 88.8 (on par with GPT-5.6 Sol), and SWEAtlas CodeBase QnA and other tests were leading or near state-of-the-art. Long-context retrieval (MRCR) was particularly strong. On the Artificial Analysis Intelligence Index, the max variant scored 62, trailing Claude Fable 5.1 and Opus 5.

The model supports a 1,000,000-token context window and is priced similarly to the previous generation (standard tier roughly $1.25 per million input tokens and $4.25 per million output tokens, with substantial discounts for caching). Meta also offers a lower-priced "contributor" tier that permits Meta to use contributed data to improve the model. The Max inference mode will be released after additional safety testing.

Technical and Market Background

The Muse Spark series is led by Meta Superintelligence Labs and accelerated after Alexandr Wang (former Scale AI founder) became Chief AI Officer. The series has been iterating rapidly since mid-2026 to close gaps with leading closed-source models. The 1.3 release coincides with Anthropic Fable 5.1 and other labs' intensive updates, reflecting intense industry competition.

Meta emphasizes the model has reached a frontier of being "almost too cheap to meter," and plans to gradually integrate capabilities into Facebook, Instagram and Meta AI products.

Potential Impact: Meta's Return to Cutting-Edge Competition and Ecosystem Positioning

1. Impact on Meta itself

Muse Spark 1.3 marks Meta's entry into the first tier on coding and agent tracks. Strong cost-effectiveness and efficiency gains should attract developers to Muse Code and the API, and lay the groundwork for upcoming personal agent products. Rapid iteration also indicates returns on the company's organizational and compute investments.

2. Impact on the AI model competitive landscape

Against OpenAI and Anthropic's continued leadership, Meta achieving parity with cost advantages may intensify price and efficiency competition. Developers will gain more high-value, lower-cost options—particularly for long-horizon coding and agent