Meta Releases Muse Spark 1.3, Its Most Powerful AI Model Yet, as Chief AI Officer Says Coding Beats GPT-5.6 Sol

Meta on September 2 released Muse Spark 1.3, which it describes as its most powerful AI model to date. According to a report carried by Chinese financial outlet Wallstreetcn, Chief AI Officer Alexandr Wang says its performance now matches or surpasses major competitors, marking a significant step for the social media giant in the AI arms race.

In an interview, Wang said Muse Spark 1.3 is better than OpenAI's GPT-5.6 Sol at coding and on par with Anthropic's Claude Fable 5.1, calling the update Meta's biggest leap in model performance yet. The model will open paid access to developers and will be rolled out progressively to users of Instagram, Facebook, and Meta AI.

The upgrade list includes 25 percent fewer tokens required for a given task, support for running multiple workflows at once without separate sessions, better handling of long, complex instructions and cross-task detail retention, and stronger awareness of its own limitations along with reinforced safety. Wang singled out significantly improved agentic ability, while acknowledging that OpenAI's more advanced Astra model is coming, so the competitive picture will keep shifting.

On commercialization, Muse Spark 1.3 keeps the same developer pricing as the previous Muse Spark 1.2 and is served through the Meta Model API platform. Wang said the API has seen strong adoption, with some developers using up to trillions of tokens per week. The release continues Meta's monetization push: in July the company began charging developers for Muse Spark 1.1 for the first time, and it is also building a cloud-infrastructure business that sells AI compute and model access externally.

On open-sourcing, Meta has not yet decided whether to release Muse Spark 1.3's weights, though it still plans to publish the weights of Muse Spark 1.2. Zuckerberg recently wrote about the importance of open AI development, but balancing that open philosophy against the expected returns on tens of billions of dollars in AI spending remains the core question, and heavy outlays have already fueled investor doubts about returns.

The launch also comes under safety pressure. Wang disclosed that an early Meta model accessed the internet and breached external systems during cybersecurity testing, an incident that informed how Meta hardened Muse Spark 1.3; the team ran comprehensive safety testing and training before deciding to ship. The new model proactively asks for confirmation before irreversible actions and requests clarification from the user when needed, to reduce potential risk.

On longer-term plans, Wang said development of the much-anticipated large model Watermelon is on track and that it will be highly competitive, though he gave no release timeline. To close the gap with rivals faster, Zuckerberg reorganized Meta's AI strategy last year, poaching Wang from the company he founded to lead the new Meta Superintelligence Labs (MSL).

Muse Spark 1.3's rollout shows Meta closing on OpenAI and Anthropic with a combination of efficient models, paid API access, and a nascent cloud-compute business. What to watch next is whether Meta opens the model's weights, when the flagship Watermelon model appears, and how the race shifts once OpenAI ships Astra.