[DigitalToday reporter Chi-gyu Hwang (황치규)] Price competitiveness is gaining increasing weight in the global race for AI leadership. Leading companies also appear to be putting much greater emphasis on price as well as performance.

Sam Altman, OpenAI CEO, unveiled the GPT-5.6 series and highlighted that it is 54 percent more token-efficient than Anthropic's latest model for agent coding tasks. SpaceXAI released its new AI model, Grok 4.5, for the first time since its listing. SpaceXAI pointed to cost efficiency as Grok 4.5's strength. It stressed that token efficiency is twice as high as other major models.

ㆍOpenAI formally launches GPT-5.6, also unveils 'ChatGPT Work' ㆍSpaceXAI releases Grok 4.5..."Opus-class model"

Meta also offered a preview of a new model, Muse Spark 1.1, focused on coding and AI agent tasks. Meta also set a low price for Muse Spark 1.1.

ㆍMeta unveils Muse Spark 1.1...challenges Anthropic and OpenAI with coding AI ㆍMeta ramps up paid AI model API sales for developers...declares price war on OpenAI and Anthropic

Calls are also growing among companies that use AI for AI costs to fall. Nikesh Arora, Palo Alto Networks CEO, drew attention by saying token costs must fall by up to 90 percent to popularise AI.

ㆍPalo Alto Networks CEO: "AI token costs must fall by up to 90 percent"

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

OpenAI will shut down its standalone desktop browser, ChatGPT Atlas, about 9 months after launch and integrate related functions into a new ChatGPT desktop app. OpenAI Deployment Company decided to acquire applied AI company Northslope. With the acquisition, OpenAI Deployment Company will expand its field-deployed engineer workforce to several hundred people to build AI systems together inside client companies.

ㆍOpenAI to end ChatGPT Atlas browser, integrate into desktop app ㆍOpenAI Deployment Company to acquire Northslope, sharply increases FDE

Anthropic expanded its AI work agent, Claude Cowork, to the web and mobile. Users can check work started at a desk on a smartphone and continue the output later even after closing a laptop.

ㆍAnthropic expands Claude Cowork to mobile and web

Microsoft has started applying its self-developed MAI model to some AI responses in Word and Excel. Microsoft is lowering its reliance on OpenAI and Anthropic software in part to cut costs and is increasing use of its own models.

ㆍMicrosoft deploys its own MAI in parts of Word and Excel...reduces reliance on external AI

Amazon Web Services, or AWS, launched a function that allows AI agents to access and run desktop applications through Amazon WorkSpaces Applications.

ㆍAWS introduces AI agent access function in Amazon WorkSpaces Applications

Cursor is reported to be developing a general-purpose AI agent that will compete with Anthropic's Claude Cowork. The move aims to diversify its business beyond coding tools. French AI startup Mistral AI unveiled its first robotics model, Robostrall Navigate. Meta released Muse Image, the first image generation model developed by Meta Superintelligence Labs. Online design platform Figma acquired the development team behind the vibe coding and AI agent platform Bird.

ㆍCursor developing general-purpose AI agent 'Sand' to compete with Claude Cowork ㆍMistral AI unveils first robotics model 'Robostrall Navigate' ㆍMeta releases image generation model Muse Image, also unveils Muse Video preview ㆍFigma acquires Bird development team...strengthens AI coding and prototyping

Tencent released the final version of its large language model, Hy3, under the Apache 2.0 licence. Hy3 is a 295 billion-parameter mixture-of-experts model with 21 billion active parameters. A preview version released in April excluded the European Union, Britain and South Korea from the licence, but the final version no longer has those restrictions.

ㆍTencent unveils final Hy3...switches to Apache 2.0 licence

Startup PrismML succeeded in running a 27 billion-parameter AI model on the iPhone 17 Pro. It is among the largest models that run on mobile devices. Active parameters for models that can run on mobile devices are usually in the billions.

ㆍPrismML succeeds in running 27 billion-parameter AI model on iPhone

LG AI Research and KOSCOM will cooperate to provide an AI-based South Korean financial data analysis service. AI-based research tech company OpenSurvey introduced a function to talk with synthetic consumers implemented with AI technology on its consumer intelligence platform Dataspace. Makinarocks will cooperate with Amazon Web Services, or AWS, as part of expanding its defence AI business. Human cloud platform Kmong created a dedicated AX team to support corporate clients' AI transformation.

ㆍLG and KOSCOM cooperate to develop an EXAONE-based 'AI service to predict the South Korean stock market' ㆍOpenSurvey introduces 'talk with synthetic consumers' function on consumer research platform ㆍMakinarocks joins hands with AWS to expand defence AI business ㆍKmong creates dedicated organisation for corporate AX...steps into corporate AX market

Wrtn Technologies is also accelerating efforts to target the corporate AX market based on its experience operating B2C services and its multi-large language model strategy. Naver Cloud signed an all-round partnership with France's Mistral AI to jointly target the manufacturing AI market.

ㆍWrtn AX expands into corporate AX using B2C experience...targets 10 billion won in revenue this year ㆍNaver Cloud and Mistral AI partner to target manufacturing

The push by AI model developers to secure their own AI chips appears to be becoming a mainstream view beyond a niche. Google, Amazon Web Services, ByteDance and Alibaba, across the United States and China, have developed customised AI chips to cut reliance on external suppliers and reduce costs. Recently, leading AI model companies are also joining the AI chip development race in earnest.

ㆍAI industry makes in-house chip development mainstream...Chinese companies also join in earnest