OpenAI recently unveiled GPT-6 Astra, which has improved at computer science problems, coding, and solving math and science questions. It can also operate a web browser and computer, write and edit code, and complete complex tasks from start to finish. It performed at the top level in mathematics, with reported progress in science, software engineering, and cybersecurity. Support for context exceeding 1 million tokens lets it handle long documents and large amounts of code at once. The model was presented as an example of generative AI competition shifting from solving difficult problems to completing complex work end to end.

Meta describes its AI agent Muse as a personalized, autonomous agent. When planning a trip, it can use email, calendars, and social media accounts the user has authorized, along with preferences inferred from earlier conversations, and proceed with bookings in a web browser. It asks for approval at steps requiring the user's judgment, such as payment, entering personal information, or sending email. It can keep tracking hotel price drops or cheaper flights, then notify the user and suggest next steps when conditions change—even after the app is closed.

Muse can create grocery lists from recipes saved on Instagram and reflect dietary preferences the user mentioned previously when suggesting meals. It is also presented as capable of handling recurring overseas business trip bookings, information searches, and schedule management. About 13 days after launch, it had recorded more than 2.5 million downloads. The summary says generative AI is moving from helping users think to continuing actions on their behalf.

This kind of personalization requires users to grant access to email, calendars, and conversation histories. Without those records, the agent may struggle to take the next step, and more sophisticated personalization can require granting it broader permissions. Even when it seeks approval for important steps, deciding how much access to sensitive personal information to allow remains a separate question. Users must set the boundaries between convenience and privacy.

Competition among AI models continues as US and Chinese companies release new models and upgrades. The summary notes that multimodal capabilities drew attention with Google Gemini after ChatGPT appeared, that more users came to prefer Claude for writing and coding, and that DeepSeek gained attention for performance and cost efficiency. It also names Moonshot AI’s Kimi and Zhipu AI’s GLM as models from Chinese tech companies. The summary says a single benchmark score cannot determine which model is best; whether a model can do the same work at lower cost also matters. Low prices are a competitive factor for Chinese models but may also weigh on profitability, while data security and regulatory concerns are cited as weaknesses.

In South Korea, projects have sought to build independent AI foundation models and a sovereign AI base by supporting GPUs, data, and talent. The summary sees practical limits to competing directly with US and Chinese companies on model size or global benchmarks, and proposes AI that is strong in Korean language and local laws and institutions, and is used in finance, healthcare, and public services. It argues that combining Korean news, publications, administrative documents, and industry data with domestic users’ lived experience could help create AI useful in South Korea, rather than simply the smartest model.

OpenAI recently unveiled GPT-6 Astra, which has improved at computer science problems, coding, and solving math and science questions. It can also operate a web browser and computer, write and edit code, and complete complex tasks from start to finish. It performed at the top level in mathematics, with reported progress in science, software engineering, and cybersecurity. Support for context exceeding 1 million tokens lets it handle long documents and large amounts of code at once. The model was presented as an example of generative AI competition shifting from solving difficult problems to completing complex work end to end.

Meta describes its AI agent Muse as a personalized, autonomous agent. When planning a trip, it can use email, calendars, and social media accounts the user has authorized, along with preferences inferred from earlier conversations, and proceed with bookings in a web browser. It asks for approval at steps requiring the user's judgment, such as payment, entering personal information, or sending email. It can keep tracking hotel price drops or cheaper flights, then notify the user and suggest next steps when conditions change—even after the app is closed.

Muse can create grocery lists from recipes saved on Instagram and reflect dietary preferences the user mentioned previously when suggesting meals. It is also presented as capable of handling recurring overseas business trip bookings, information searches, and schedule management. About 13 days after launch, it had recorded more than 2.5 million downloads. The summary says generative AI is moving from helping users think to continuing actions on their behalf.

This kind of personalization requires users to grant access to email, calendars, and conversation histories. Without those records, the agent may struggle to take the next step, and more sophisticated personalization can require granting it broader permissions. Even when it seeks approval for important steps, deciding how much access to sensitive personal information to allow remains a separate question. Users must set the boundaries between convenience and privacy.

Competition among AI models continues as US and Chinese companies release new models and upgrades. The summary notes that multimodal capabilities drew attention with Google Gemini after ChatGPT appeared, that more users came to prefer Claude for writing and coding, and that DeepSeek gained attention for performance and cost efficiency. It also names Moonshot AI’s Kimi and Zhipu AI’s GLM as models from Chinese tech companies. The summary says a single benchmark score cannot determine which model is best; whether a model can do the same work at lower cost also matters. Low prices are a competitive factor for Chinese models but may also weigh on profitability, while data security and regulatory concerns are cited as weaknesses.

In South Korea, projects have sought to build independent AI foundation models and a sovereign AI base by supporting GPUs, data, and talent. The summary sees practical limits to competing directly with US and Chinese companies on model size or global benchmarks, and proposes AI that is strong in Korean language and local laws and institutions, and is used in finance, healthcare, and public services. It argues that combining Korean news, publications, administrative documents, and industry data with domestic users’ lived experience could help create AI useful in South Korea, rather than simply the smartest model.