AI Daily Newsletter: GPT-5.6 Sol Ultra solves 50-year-old math problem and Apple takes on OpenAI

Today’s AI Daily newsletter (12/07/2026) brings you all the hottest news that’s just been released. From OpenAI’s super AI solving a 50-year-old math problem by running in parallel, to the intense legal battle between Apple and OpenAI. We’ll also take a look at China’s top-notch Orca world model and Meta Muse Spark 1.1’s impressive cost breakthrough. Let’s review the 5 most notable AI news of the day!

🧮 GPT-5.6 Sol Ultra solves a 50-year-old math problem in under an hour

It’s true that when AI learns math, humans can only be left speechless. OpenAI’s GPT-5.6 Sol Ultra has successfully proven the Cycle Double Cover Conjecture - a math problem that has been puzzling mathematicians for half a century. What’s astonishing is that this model took less than an hour to solve the problem by mobilizing 64 sub-agents to run in parallel. The math community is currently reviewing the lengthy proof document to see if there are any errors, but this is indeed a remarkable milestone that showcases the power of Agentic workflow.

Source: The Decoder

🍏 Apple officially sues OpenAI for stealing trade secrets

The tech world drama has been heated up again as Apple decides to take OpenAI to court for allegedly stealing trade secrets. Apple claims that the theft was directly ordered by OpenAI’s top executives, with the help of a former long-time Apple employee who has defected to the rival company. It seems that the AI race is not only intense on the benchmark charts but also extremely fierce in the courtroom.

Source: TechCrunch AI

🐋 BAAI China launches Orca: A self-learning world model from 125,000 hours of video

The Beijing Academy of Artificial Intelligence (BAAI) has introduced a world model called Orca, which can predict abstract world states instead of just predicting tokens or pixels. Interestingly, this “killer whale” was trained on 125,000 hours of video without any action labels, yet it still achieves performance comparable to specialized robotics systems. This is a significant step forward, showing that the direction of training world models without labels is becoming increasingly effective.

Source: The Decoder

⚡ Meta Muse Spark 1.1: Surpasses GLM-5.2 in coding while being cheaper

Meta continues to affirm its position in the open-source model race by upgrading Muse Spark to version 1.1, which has increased its score by 8 points on the Artificial Analysis evaluation index in just three months. Notably, it has surpassed GLM-5.2 in programming with an impressive score of 71.3, while the operating cost is only around $0.26 per task. It’s true that in the AI competition era, the model that is both good and cheap will be the choice of programmers.

Source: The Decoder

📸 Meta backpedals: Removes controversial AI feature on Instagram

After a wave of intense backlash from the user community, Meta has hastily removed a new AI feature on Instagram. The feature automatically scanned and referenced public posts from users to create new content, making netizens feel that their privacy was severely compromised. Although Meta claims that they only wanted to provide a useful creative tool, it’s clear that forcing users to “donate” their data has never been a wise move.

Source: TechCrunch AI