From GPT-6 Astra to AI Factory, This Week the AI Landscape Completely Changed

▲ Photo: AI-generated image

In the first week of September, OpenAI, Anthropic, and Google DeepMind successively unveiled new models within three days, redrawing the competitive landscape of the AI industry.

OpenAI announced its next-generation model, GPT-6 Astra, on the 3rd (local time). The company introduced this model as "the most capable model optimized for the most challenging end-to-end tasks." It is characterized by its ability to weave together various tools, such as web browsing, code execution, and computer operation, into a single workflow, and supports a context window of over 1 million tokens.

OpenAI executives stated in the model unveiling briefing that they invested more computing resources than ever before in safety reviews, and that this model underwent official review procedures with the US administration before its official launch.

Two days earlier, on the 1st, Anthropic released Claude Fable 5.1 and Mythos 5.1, reducing cache read fees by 75%. Then, on the 2nd, Google DeepMind unveiled Gemini 3.8 Flash. With the announcements from the three companies concentrated within three days, some in the industry describe this week as one where the three major model developers clashed simultaneously.

In Korea, Naver announced on the 7th its plan to build a large-scale AI factory in partnership with Nvidia and Brookfield. The plan is to start operations in 2027 with a 55-megawatt capacity, expanding to a 1-gigawatt facility in the future, which is approximately four times the capacity of Naver's existing largest data center, 'Gak Sejong'. Kakao stated that it is continuing to advance its conversational agent within KakaoTalk.

However, the developer community's reactions to the rush of new models are mixed. Many evaluations on benchmark comparison sites and online communities indicate that Astra showed clear improvements over previous models in computer usage and coding, but voices also emerged pointing out that details of parameters and architecture were not disclosed.