Meta has released "the strongest model ever," Muse Spark 1.3, which is aligned with GPT-5.6 Sol, further intensifying the competition among leading players in the AI field.

On Wednesday local time, Meta Platforms (META.US) released its most powerful artificial intelligence (AI) model to date, Muse Spark 1.3. Meta's Chief AI Officer, Alexandr Wang, stated that this represents the company's "largest leap" in model performance so far, particularly in coding and agent tasks, bringing Meta closer to leading competitors like OpenAI and Anthropic. Developers can access and pay for Muse Spark 1.3 through the Meta Model API starting Wednesday. Meta also noted that it will soon roll out this update to users of its social media products, including Instagram and Facebook, as well as Meta AI. "Significant performance leap" targeting coding and agent tasks In an interview on Wednesday, Wang stated that Muse Spark 1.3 has made noticeable advancements in coding and agent capabilities, with agent capabilities referring to the model's ability to complete multiple tasks on behalf of human users. He mentioned that this update positions Meta on par with models recently released by OpenAI and Anthropic. Wang further commented that Muse Spark 1.3 is "competitive" with Anthropic's Claude Fable 5.1 and is "superior" in coding tasks to OpenAI's GPT-5.6 Solhowever, OpenAI is set to release a more advanced model, Astra. Wang also believes that Meta's new model "outperforms any Chinese model currently on the market." However, it is important to note that direct comparisons between models are not straightforward. Different models have their strengths and weaknesses across various tasks, and benchmark parameters may be deliberately optimized, which may not fully reflect real-world usage scenarios. From the benchmark data released by Meta, Muse Spark 1.3 scored 75.4 on DeepSWE v1.1, outpacing GPT-5.6 Sol and Opus 5 listed in the chart. Compared to the previous generation Muse Spark 1.2, the model also saw significant improvements in long context tests. Nonetheless, Meta acknowledged that in several other benchmarks, competitors' models still maintain an edge. Efficiency improvements: reduced token consumption, parallel processing of multiple tasks Compared to version 1.2, Muse Spark 1.3 shows significant improvements in efficiency. Meta engineers found that the number of tool calls required by the new model has decreased by approximately 20%, and token consumption has dropped by about 25%. This means the model has fewer "detours" in day-to-day workloads, with responses becoming more concise and reducing unnecessary interaction rounds. This efficiency boost is especially important for coding tasks. Long tasks require the model to continuously remember the user's initial requirements and execute accordingly. Muse Spark 1.3 can handle multiple workflows simultaneously within the same conversation without needing to open multiple independent sessions; it can also better manage long and complex instructions, retaining key details across multiple tasks. Meta also indicated that the new model has a stronger awareness of its own limitations, being more willing to proactively ask users for clarification when requests are vague, rather than proceeding based on incorrect assumptions. Before performing irreversible actions, the model will seek confirmation, thereby reducing the risk of operational errors. Additionally, Wang revealed that Meta conducted comprehensive safety testing and security training before deciding to release Muse Spark 1.3. It is reported that an earlier version of Meta's model had autonomously accessed the internet during cybersecurity tests, invading external service systems. This incident resonates with recent similar events involving other model companies, raising concerns about the controllability of AI technology. Wang stated that this incident helped Meta improve the safety and protection strategies for Muse Spark 1.3. Pricing remains unchanged, continuing to bet on the developer ecosystem In terms of commercialization, Meta has chosen to keep the API pricing for Muse Spark 1.3 unchanged. The standard pricing is set at $1.25 per million input tokens, $0.15 per million cached input tokens, and $4.25 per million output tokens. Similar to the previous version, Meta continues to offer a cheaper "contributor" tier for developers who allow their data to be used for model improvement, priced at $0.10 per million input tokens and $0.20 per million output tokens. Wang noted that the adoption rate of the Meta Model API platform has been strong, with some developers "using tens of trillions of tokens every week." This marks Meta's ongoing attempts at AI commercialization. In July of this year, Meta began charging developers for the use of Muse Spark 1.1. At that time, Zuckerberg expressed his desire for Meta's model to become one of the most affordable options on the market. Regarding the open-source strategy, although Zuckerberg has recently emphasized the importance of open AI development, Meta has yet to decide whether to open-source the weights of Muse Spark 1.3. The weights are internal parameters formed during the model training process that determine how the model responds. If open-sourced, external developers can download, run, or further develop based on the model. Wang indicated that Meta still plans to release the weights of the previous version, Muse Spark 1.2. A final decision regarding the open-sourcing of 1.3 has not yet been made. Meanwhile, Meta is continuing to advance its larger and highly anticipated model, Watermelon. Wang stated, "We believe Watermelon will be highly competitive." However, he declined to reveal a specific release date. Currently, Meta is investing hundreds of billions of dollars in an attempt to catch up with the leaders in the rapidly changing AI race. Last year, Zuckerberg adjusted the company strategy, bringing Wang from the company he founded to Meta and appointing him to lead the newly established Meta Superintelligence Labs (MSL). Since then, Wang has been pushing for a faster pace in releasing new models and products to bridge the technological gap with competitors. However, Meta's massive expenditures have also drawn investors' attention, as the market is hoping to see clearer returns on investment. Against this backdrop, Meta has begun to lay the groundwork for its cloud infrastructure business, selling AI computing power and model access.