AI News Flash

Anthropic reaches $1.5 billion copyright settlement with authors

Anthropic has agreed to a $1.5 billion settlement with a group of authors to resolve a class-action lawsuit over the alleged use of pirated books to train its AI models. The agreement highlights the growing legal and financial risk surrounding how AI companies obtain and use training data.

For companies building or fine-tuning AI systems, copyright compliance is becoming a core operational requirement. Teams need documented data provenance, clear licensing terms, and review processes that can identify unauthorized material before it enters a training pipeline.

Google develops Frozenv2 chip to make Gemini AI dramatically more efficient

Google is developing a new AI chip code-named Frozenv2 to improve the efficiency of running Gemini models. The chip is expected to process six to ten times more tokens per unit of power than current hardware and is reportedly planned for deployment in 2028.

Large gains in tokens per watt could lower inference costs and accelerate the commercial rollout of large models. The impact may be especially important for cloud computing, where power and capacity constrain deployment, and for edge systems that must deliver useful AI performance within tighter energy limits.

Kimi pauses new consumer subscriptions as compute capacity runs tight

Moonshot AI has temporarily stopped accepting new consumer subscriptions to Kimi because its available compute capacity is under pressure. The company is directing its current resources toward existing subscribers so their benefits and service quality remain unaffected, while expanding capacity as quickly as possible. New subscription slots will reopen gradually as additional compute comes online.

Kimi also plans to separate its main membership, covering Kimi Web, the Kimi app, and Kimi Work, from Kimi Code benefits. The split is intended to match compute more precisely to different workloads and keep the experience stable as demand grows.

Alibaba previews the 2.4T-parameter Qwen3.8 and says an open release is coming

Alibaba says its next-generation Qwen3.8 model is approaching release and is planned for an open release. A 2.4-trillion-parameter Qwen3.8-Max-Preview is already available through Alibaba's Token Plan, Qoder, and QoderWork, giving developers an early look before the full launch.

Alibaba reports substantial gains over Qwen3.7-Max on complex engineering and knowledge-work tasks, including full-stack development, data analysis, office workflows, multi-agent coordination, and long-running assignments. If the open release follows, it could lower the barrier for developers and smaller companies that want to build on a frontier-scale model.

Apollo's chief economist says the new 60/40 portfolio is AI versus non-AI

For decades, the classic 60/40 portfolio meant putting 60% into stocks for growth and 40% into bonds for stability. Apollo Global Management chief economist Torsten Slok argues that today's more relevant divide is between AI and non-AI exposure. His point is less about prescribing a literal allocation and more about recognizing that the AI boom now touches so much of the market that avoiding it entirely has become increasingly difficult for investors.

SpaceXAI open-sources Grok Build and resets usage limits

Elon Musk's SpaceXAI has released Grok Build as open-source software and reset usage limits for all users. The coding agent is now available for developers to inspect and run with greater control over how it fits into their workflows.

Addressing privacy concerns, the company says Grok Build has honored zero data retention settings since launch and has always allowed users to disable data uploads from the command-line interface. During the early beta, data retention was enabled by default for users who had not selected zero data retention. SpaceXAI says it disabled default retention for every Grok Build user on July 12 and is deleting previously retained coding data.

MiniMax Code 2.0 targets smoother long-running tasks, with remote and browser control coming soon

MiniMax has released the desktop version of MiniMax Code 2.0, with a focus on reducing interruptions, improving context continuity, and making the waiting periods in long-running tasks easier to manage. Later this month, the company plans to add remote control, browser control, goal mode, and plan mode, expanding the product's ability to handle complex tasks and coordinate work across different environments.

Anthropic reportedly begins investor meetings ahead of a possible October IPO

Anthropic has reportedly begun arranging meetings between its management team and prospective investors as it prepares for an initial public offering. Goldman Sachs, Morgan Stanley, and JPMorgan are said to be serving as underwriters, with a U.S. listing possible as early as October.

The Claude maker confidentially filed IPO documents with the U.S. Securities and Exchange Commission in June. According to the report, Anthropic was valued at $96.5 billion after a May funding round, above OpenAI's reported valuation of about $85.2 billion at the time.

Alibaba launches Qwen-Audio-3.0-Realtime for natural, full-duplex voice conversations

Alibaba has introduced Qwen-Audio-3.0-Realtime, a real-time speech interaction model built for faster, more natural conversations. The release brings upgrades across four areas: general intelligence, agent tool calling, empathetic dialogue, and the fluency of full-duplex interaction. The goal is to let people and AI speak with less turn-taking friction while giving voice agents stronger reasoning and practical action-taking abilities.

ModelBest raises more than RMB 5 billion as its valuation tops RMB 20 billion

Chinese AI company ModelBest has completed a new funding round backed by national funds, centrally administered state-owned enterprises, automakers, industry investors, and financial institutions. Its total financing in the first half of 2026 now exceeds RMB 5 billion, while its valuation has climbed above RMB 20 billion.

Founded in 2022 as a spinout from Tsinghua University's Natural Language Processing Lab, ModelBest focuses on bringing native on-device large models into real products. Its technology has already been deployed across smartphones, vehicles, aviation, and embodied intelligence, with further expansion planned for industrial, household, and specialized applications.

ChatGPT returns to normal after a brief service disruption

OpenAI says ChatGPT has fully recovered from a service disruption that caused login problems and intermittent errors. The company applied a mitigation, monitored the recovery, and confirmed that all affected services were back to normal roughly 42 minutes after the incident began.

OpenAI reports login issues and intermittent errors affecting ChatGPT

OpenAI reported a partial ChatGPT outage after users encountered login issues and intermittent errors. The company began investigating the incident before applying a mitigation and monitoring the platform's recovery. A later update confirmed that the disruption had been resolved.

OpenAI reportedly plans a screenless AI companion speaker as its first hardware product

OpenAI is reportedly developing a movable, screen-free smart speaker designed to serve as a new kind of home computer for the AI era. Positioned as a home AI companion with a more human-like presence, the device could control smart-home products, play media, answer questions, reply to messages, and bring ChatGPT's broader capabilities into a dedicated physical product.

The project faces legal uncertainty after Apple sued OpenAI last week, alleging theft of trade secrets. OpenAI argues that the planned device is meaningfully different from anything Apple currently sells and is unlikely to infringe the iPhone maker's trade secrets.

OpenAI's hardware group is said to be working on roughly five products, with the speaker expected to come first as a physical embodiment of ChatGPT. The company reportedly hopes to announce it this year and launch it in 2027, although the litigation could affect that timetable.

OpenAI seeks more than $1 million in legal fees from xAI

OpenAI has asked a federal judge to order Elon Musk's xAI to pay more than $1 million in legal expenses after xAI's trade-secret lawsuit against the ChatGPT maker was dismissed. OpenAI argues that the case should never have been filed and says defending against it unnecessarily consumed company resources.

U.S. stocks open mixed as AI and memory shares retreat

Major U.S. stock indexes opened in different directions, with the Dow Jones Industrial Average edging up 0.06%, the S&P 500 falling 0.31%, and the Nasdaq Composite dropping 0.70%. AI-related technology stocks pulled back broadly, while memory and storage names led the declines: SK hynix ADRs fell 9%, Micron Technology lost 5%, and SanDisk dropped 6%.

StepFun unveils an AI-native agentic smartphone

StepFun has unveiled an Agentic Phone designed natively around a large AI model. The device runs Step AOS, an agent-native operating system, and includes the Amoo AI agent. StepFun describes it as a new category that brings the model, software, and hardware together in a single product. It is scheduled to make its public debut at the World Artificial Intelligence Conference, which opens on July 17. Detailed specifications, pricing, and availability have not yet been announced.

Shengyi Electronics expects first-half net profit to more than double

Shengyi Electronics expects net profit attributable to shareholders to reach RMB 1.082 billion–1.137 billion in the first half of 2026, an increase of 104%–114% from a year earlier. The company said strong demand from AI computing and high-speed communications, a stronger focus on premium markets, an improved product mix, and contributions from its intelligent computing center project drove the gains.

TCL Technology forecasts a 96%–108% jump in first-half net profit

TCL Technology expects first-half net profit attributable to shareholders to reach RMB 3.70 billion–3.92 billion, up 96%–108% year over year. The company attributed the increase to significantly stronger profitability in its semiconductor display business and continued improvement in its solar photovoltaic operations.

Samsung and SK hynix sell-off triggers South Korea's seventh circuit breaker of the year

A sharp sell-off in major technology stocks, including Samsung Electronics and SK hynix, sent South Korean shares tumbling and triggered a market circuit breaker for the seventh time this year. The benchmark index closed nearly 9% lower, while Samsung fell 10.7% and SK hynix dropped 15.37%. SK hynix is now down almost 40% from its June peak.