TranslateGemma shows how targeted training helps Google squeeze more performance out of smaller models: the 12B version translates better than a base model twice its size and runs on a regular laptop. With the growing Gemma family, Google is staking its claim in the race for open AI models.
Apple is betting billions on Google to power Siri’s AI future, and OpenAI says it walked away from the deal on purpose.
What China can do, the US can do too: OpenAI has published a call for proposals to boost domestic AI hardware production. AI relies on a broad ecosystem of physical components beyond chips, OpenAI says. The company is seeking manufacturers and suppliers of data center components like cooling systems, power supplies, and networking equipment, plus consumer electronics and robotics. Applications are open through June 2026.
The move comes as China reportedly restricts Nvidia H200 imports and pushes domestic manufacturers to source hardware locally. If China succeeds in decoupling its supply chain, the US can expect countermeasures—and if both countries want true independence, they'll each have to build their own.
The initiative fits squarely with Trump's "America First" agenda. OpenAI frames it as "reindustrialization of the country." Notably, OpenAI President Greg Brockman donated $25 million to Trump's campaign.
OpenAI has invested in Merge Labs, a startup developing brain-computer interfaces (BCIs). The company raised a total of $252 million in seed funding, according to Bloomberg. OpenAI CEO Sam Altman is a co-founder of Merge Labs. The exact size of OpenAI's investment remains unknown.
Merge Labs aims to combine biological and artificial intelligence to improve human capabilities. The startup is working on safer BCIs with higher bandwidth by combining biology, devices, and AI, according to OpenAI. OpenAI views BCIs as an important way for humans to interact more naturally with AI systems. The collaboration includes basic science models and tools to speed up research.
Merge Labs relies on ultrasound technology instead of electrodes. The goal is to create less invasive BCIs that can interact with more neurons. This puts OpenAI in direct competition with Elon Musk's Neuralink, which was founded in 2016 and uses electrode-based implants. The investment follows OpenAI's strategy of backing hardware and interface technologies. Rumors about this investment had circulated since summer.
OpenAI has launched a standalone translation tool built on ChatGPT. "ChatGPT Translate" supports more than 25 languages with an interface similar to Google Translate or DeepL—two text fields with automatic language detection for the input.
Users can refine translations with additional prompts, for example, switching to a business tone or simplifying for children. These prompts redirect to the main ChatGPT interface, suggesting the tool is mainly designed as a gateway to the chatbot. OpenAI hasn't officially announced it yet.
Unlike full ChatGPT, the tool only handles text with apparent length limits. During testing, it sometimes returned chatbot responses asking for clarification instead of translations, with no way to respond. This suggests it's essentially a prompt in a new interface rather than a specialized translation model. For now, ChatGPT itself remains the more capable option.
The US AI industry continues to deliver drama. OpenAI has rehired three former employees from Mira Murati's AI startup Thinking Machines, including co-founder Barret Zoph, who was recently fired as CTO. Thinking Machines let Zoph go for "unethical conduct," according to journalist Kylie Robison, citing two sources.
A source speaking to Wired claims Zoph passed confidential company information to competitors. OpenAI, the company he's now returning to, would be the obvious candidate. OpenAI's product CEO Fidji Simo announced Zoph's return in an internal memo and on X, saying the company doesn't share Murati's concerns. According to Simo, his return had been "several weeks" in the making.
Zoph and Metz originally worked at OpenAI before co-founding Thinking Machines with Murati, who served as OpenAI's CTO before starting the company. Along with Zoph and Metz, Sam Schoenholz is also returning to OpenAI. Zoph will report directly to Simo. Soumith Chintala takes over as CTO at Thinking Machines.
OpenAI has released GPT-5.2 Codex to developers through the Responses API. The model was previously limited to the Codex environment. According to OpenAI Developers, it excels at complex, tedious tasks like developing new features, refactoring code, and tracking down bugs. OpenAI also says it's their best cybersecurity model yet, helping identify vulnerabilities in codebases.
The model accepts text and images as input and offers four levels of reasoning effort: low, medium, high, and very high. Pricing comes in at $1.75 per million input tokens and $14 per million output tokens, a notable increase from earlier GPT-5 Codex models, which cost $1.25 per million input tokens and $10 per million output tokens.
Coding platforms Cursor and Windsurf have already integrated the model, with Windsurf offering it at half price for a limited time. OpenAI has published a prompting guide.
Microsoft is shaping up to be one of Anthropic's best customers. According to The Information, the software giant is on track to spend nearly $500 million a year on Anthropic's models. Microsoft is using the OpenAI competitor's AI technology in products like GitHub Copilot and Office applications, likely in part to strengthen its bargaining position with OpenAI.
Microsoft has also told its Azure sales team to count Anthropic model sales to cloud customers toward their quotas, just like they would for Microsoft's own software. That's unusual for third-party products, which typically generate less revenue for Azure. The deeper collaboration follows Microsoft's investment of up to $5 billion in Anthropic last November.
