Global competition in AI is being “reset” as Chinese AI companies step up their offensive. While the U.S. government regulates who OpenAI and Anthropic can make their latest high-performance models available to, Chinese companies are rapidly expanding global market share with open-source models that anyone can download and modify, and that are also cheaper.

ㆍIn the global AI market, a strong Chinese AI wave armed with cost-performance is intensifying ㆍChina does it again in AI… “GLM-5.2” matches GPT and cuts costs by 48 percent, surprising the U.S. ㆍHow fast is Chinese AI catching up with the U.S.… opinions differ in the tech sector ㆍByteDance releases “Seedance 2.5” AI video generation model… creates videos up to 30 seconds

Chinese AI companies are also being assessed as showing performance comparable to Anthropic’s top model, Mythos, in some cybersecurity scenarios.

ㆍ“Chinese AI has Mythos-class power in security”… backlash to Trump AI policy ㆍ“Chinese open-source AI GLM-5.2 lowers the barrier to cyberattacks”… security researchers warn ㆍChina’s 360 Security says it developed a Mythos rival… unveils vulnerability-detection AI

The U.S. government appears to be easing overseas export bans it took on Mythos5 and Fable5 for Anthropic. Reports also say the Trump administration could lift access restrictions on Anthropic’s Fable5 model as early as within 1 week. Even so, a national stance on export controls appears to remain in place.

ㆍU.S. Commerce Department allows limited provision of Anthropic Mythos5… Fable5 still restricted ㆍRestrictions on access to Anthropic Fable5 to be lifted soon… talks between Anthropic and U.S. government make rapid progress ㆍOpenAI limits initial access to a new AI model, at government request, to “trusted partners”

Meanwhile, Japanese AI startup SakanaAI has launched an AI system called Fugu that does not rely on a single model, but orchestrates multiple AI models through a single API to handle complex tasks. SakanaAI said Fugu Ultra delivered performance on par with Anthropic’s Fable5 and Mythos Preview on core engineering, science and reasoning benchmarks, and outperformed Fable5 in some items.

ㆍJapan’s Sakana launches multi-model AI “Fugu”… “beats Claude Fable5 on some benchmarks”

The article also summarizes moves by companies in and outside South Korea around AI.

OpenAI jointly developed the AI inference server chip Jalapeno in cooperation with Broadcom. It is part of OpenAI’s strategy to reduce dependence on Nvidia chips and secure its own hardware.

ㆍOpenAI unveils its own AI inference server chip… first operation by year-end

Anthropic launched an always-on AI team member, “Claude Tag,” as a research preview in Slack. Claude Tag lets users tag @Claude in Slack channels to get insights or assign tasks. Anthropic is preparing to support mobile environments with Claude Cowork, its desktop-centered agentic mode.

ㆍAnthropic adds AI teammate “Claude Tag” to Slack… always-on ㆍAnthropic preparing mobile support for Claude Cowork

Google is restructuring the Strike team it recently launched to chase Anthropic in the competition for AI coding tools. Google’s strategy is to split roles so the mid-training team strengthens model capabilities, including coding, while the post-training team focuses on user experience, to speed up improvements. Salesforce released a customer service AI agent called “Help Agent” and introduced a new pricing system that charges only when a problem is actually resolved.

ㆍGoogle revamps AI coding Strike team… launches mid-training team ㆍSalesforce unveils AI agent for customer support… charges only when problems are resolved

Microsoft added 3 features for financial professionals to Excel Copilot. They are skills, plans and external data integration. Productivity software company Notion will end its email product Notion Mail on Sept. 22. It is part of a shift to focus on AI agent services. AI coding tool Cursor acquired open-source coding assistant Continue.

ㆍMicrosoft adds finance skill and plan features to Excel Copilot… also links external data ㆍNotion to end email app Notion Mail in September… focus on AI agents ㆍCursor acquires open-source coding assistant Continue

Amazon Web Services (AWS) unveiled in preview a new “continuous modernization” feature for AWS Transform that analyzes and resolves technical debt across large code repositories. Nvidia unveiled a safety system called “Nvidia Halo for Robotics,” designed to allow humanoid robots to be deployed alongside humans. Enterprise AI platform company Dataiku unveiled “Cobuild,” an AI project auto-generation tool.

ㆍAWS Transform upgrade… adds autonomous detection and resolution of code technical debt ㆍNvidia unveils humanoid robot safety system “Halo” ㆍDataiku says performance matters for AI projects too… automation tool “Cobuild” is the answer

Celts, which develops search infrastructure for AI agents, raised $12.5 million in seed funding. Celts plans to use the funds to expand search infrastructure that supports AI agents in finding information on the web. AI agent memory startup Aristo (Aristo) received pre-seed investment from Kakao Ventures.

ㆍCelts raises $12.5 million… develops search infrastructure for AI agents ㆍAI agent memory startup Aristo raises pre-seed funding from Kakao Ventures

KT Cloud will build an AI safety cooperation framework with Kakao to implement safe AI services. LG CNS launched “Perfectwin ERP Edition,” an ERP test automation solution based on real transaction data specialized for SAP ERP. SK Telecom will fully apply its self-developed proprietary AI foundation model to manufacturing sites. AI research tech company OpenSurvey is revamping its consumer intelligence platform Dataspace to an AI agent-based system to support everything from survey planning to writing result reports without specialized knowledge. NC AI completed development of a next-generation model for its 3D generative AI service, “VARCO 3D.” It plans to apply it to its own cloud-based software-as-a-service platform in July.

ㆍKT Cloud-Kakao cooperate on AI stability ㆍLG CNS launches ERP test solution with agentic AI ㆍSKT begins full deployment of proprietary AI model at manufacturing sites ㆍOpenSurvey revamps enterprise research platform based on AI agents ㆍNC AI unveils “VARCO 3D 2.0”… service application in July

Lee Hak-joon (이학준), CEO of Madras Check, said at the recent “AX Festa 2026” that “the era of collaboration platforms used only by people is over,” and declared it would transform its collaboration tool Flow into an AI work agent platform.

ㆍFlow to shift to AI work agent… “era of people-only collaboration is over”

As AI use rises, costs are burdensome, driving active efforts by companies to use AI as cost-effectively as possible. More companies are trying to cut AI costs and to mix cheaper models as needed in addition to high-performance models from OpenAI and Anthropic. In this process, router software that can automatically select cheaper models depending on task complexity is drawing attention. Some companies are also adopting open-source models and customizing them to cut costs.

ㆍ[Tech Inside] “Cost-performance is also competitiveness”… evolving technology for cutting AI costs ㆍ“By 2028, AI coding costs will exceed developers’ average annual salary”… Gartner outlook