The AI startup from former OpenAI CTO Mira Murati had a rough week. After firing a co-founder, several researchers left for their former employer. A planned funding round at a $50 billion valuation could be in jeopardy.
South Koreans now spend more per month on AI subscriptions than on Netflix. According to Hankyung Aicel, payments for seven AI services, including ChatGPT and Gemini, hit an estimated 80.3 billion won (roughly $55-60 million) in December 2025. That's more than the average monthly Netflix subscription revenue in Korea during 2024, which came in at 75 billion won (around $50-55 million). One important caveat: the AI figure includes business payments, while Netflix is a consumer-only service.
Credit card payments for AI services jumped from 52,000 transactions in January 2024 to 1,666 million in December 2025. Private customers paid an average of 34,700 won (about $24), while businesses spent 107,400 won (roughly $74). ChatGPT dominated with 71.5 percent of all payments, followed by Gemini at 11.0 percent and Claude at 10.7 percent. According to Hankyung Aicel CEO Kim Hyung-min, Korea's subscription market continues to grow, and generative AI is becoming a regular subscription product.
For context: Netflix reports revenue per subscription of around $7 for Asia-Pacific, compared to roughly $17 in the US and Canada. That's significantly higher revenue per subscription per month.
Google is rolling out Gemini 3 Pro to power AI Overviews in search. The system now automatically routes complex queries to Google's most powerful language model, while faster models still handle simpler questions, according to Robby Stein, product manager for Google Search.
This intelligent routing already works in AI Mode, Google's AI-powered search chat, and is now expanding to AI Overviews, the quick answers that appear directly below search queries. The feature is available worldwide in English, but only for paying Google AI Pro and Ultra subscribers.
AI Overviews and similar services from other companies have faced criticism for confidently delivering incorrect answers. While source citations create an appearance of trustworthiness, users rarely verify them. More capable models can reduce errors but won't eliminate them.
Chinese AI startup DeepSeek ran into trouble developing its new flagship model and had to switch to Nvidia chips. According to insiders, Deepseek initially tried using chips from Huawei and other Chinese manufacturers last year, reports the Wall Street Journal. But the results weren't good enough. The company ended up switching to allegedly smuggled Nvidia chips for some training tasks, which finally got things moving. The new model is expected to ship in the coming weeks.
At a recent conference in Beijing, leading Chinese AI researchers admitted that Chinese AI models won't be able to keep pace with US companies without access to better hardware. Justin Lin from Alibaba's Qwen team put the odds of overtaking OpenAI or Anthropic within three to five years at 20 percent at best. Meanwhile, the Chinese government is pushing to cut US chip imports to boost domestic production.
OpenAI is pushing "Open Responses," an open interface that works with language models from different providers. The project builds on OpenAI's Responses API and lets developers write code once and run it with any AI model.
Currently, Google, Anthropic, and Meta all handle their APIs differently, which means developers have to rewrite code when switching between models. Open Responses tries to fix that with a shared format for requests, responses, streaming, and tool calls. Vercel, Hugging Face, LM Studio, Ollama, and vLLM have already signed on.
Of course, if successful, this move works in OpenAI's favor. If its API becomes the default, competitors would need to adapt to OpenAI's approach, while existing OpenAI customers wouldn't have to change a thing. The "open" label also lets the company signal a spirit of collaboration, even though it's not sharing any technology beyond what's already public.
OpenAI will start testing ads in ChatGPT, despite CEO Sam Altman’s earlier objections. With a valuation of up to $750 billion to justify and only around five percent of users paying for the service, the company is under enormous pressure to find new revenue streams.
Anthropic has expanded access to its new Claude Cowork feature. When Cowork debuted on Monday, it was limited to Max subscribers, a $200 per month tier that put it out of reach for most users. Pro subscribers can now access the feature for $20 per month, though Anthropic warns they may hit usage limits faster since Cowork consumes more tokens than regular chat. Max subscribers still get higher usage limits.
Cowork brings the agent-based capabilities of Claude Code to the desktop app for everyday tasks that don't require programming knowledge. With computer access enabled, Claude can handle more complex tasks on its own: sorting files, gathering context from multiple documents, and similar workflows. For now, the feature remains exclusive to the macOS desktop app.
Anthropic has already shipped several updates since Monday's launch: users can now rename sessions, connections to external services are more reliable, file previews work better, and the app prompts for confirmation before deleting files.
Cloudflare is acquiring British startup Human Native to create a new payment system for AI training data. The company runs a marketplace for AI training data and converts multimedia content into structured, licensable datasets.
The move addresses a growing problem: AI crawlers scrape the web on a massive scale without paying website operators. Instead of letting AI companies grab content for free, publishers could make their data available through an index and get paid for it.
Cloudflare has already been developing tools like "AI Crawl Control" and "Pay Per Crawl" that let website operators control who can use their content for AI training. The company also co-founded the x402 Foundation with Coinbase to enable automatic machine-to-machine payments. Cloudflare runs its own AI platform and recently expanded it by acquiring Replicate.
