SpaceXAI's Grok 4.5 goes 'Opus-class' + OpenAI's GPT-Live voice

SpaceXAI released Grok 4.5 to the public, its new flagship built on the 1.5-trillion-parameter V9 foundation and trained alongside Cursor, the coding editor SpaceX agreed to acquire for $60 billion in June. Elon Musk calls it "Opus-class" — comparable to Anthropic's Claude Opus 4.8, but faster and cheaper at $2 per million input tokens and $6 per million output. On SWE-Bench Pro, xAI reports it resolves tasks in roughly a quarter of the tokens Opus uses, though it trails Opus on two of the four coding benchmarks xAI itself published.

OpenAI began replacing ChatGPT's Advanced Voice Mode with GPT-Live, a new pair of full-duplex voice models — GPT-Live-1 and a smaller GPT-Live-1 mini — that can listen and speak at the same time. The models handle turn-taking more naturally, interjecting "mhmm" or staying quiet, and hand off to a frontier model behind the scenes for questions that need web search or deeper reasoning. Free users get the mini model by default; paid tiers get the larger one.

The Swedish vibe-coding startup Lovable is reportedly in talks to raise $300 million at a $13.2 billion valuation, exactly double the $6.6 billion it hit last December. Lovable — whose tools let people build software by describing it in plain language — reached a $500 million annualized revenue run rate in June, less than three years after launching. Menlo Ventures is expected to lead the round, underscoring how vibe coding has become the most lucrative use case in consumer AI.

After a December campus shooting, Brown University economics professor Roberto Serrano offered a take-home midterm for the first time in nearly two decades — and 40 of 86 students scored a perfect 100, lifting the class average to 96, versus a historical range of 65 to 80, on an exam that was harder than usual. Suspecting ChatGPT, Serrano moved the final in person; the average crashed to 48. The episode turns a fuzzy worry about AI cheating into a hard number, and Serrano says Brown's response has been "meek."

When a fabricated image of Senator Mitch McConnell in a hospital bed spread across Reddit and X this week, fact-checkers at Snopes debunked it by detecting Google's SynthID watermark — an invisible signature baked into AI-generated images that survives screenshots, resizing, and compression. It was the watermarking system's first high-profile save. The catch: SynthID only works when the image generator participates, and while Gemini has embedded it since 2025 and OpenAI joined in May, most tools still don't.

A class-action lawsuit against xAI was expanded this week to add two more plaintiffs who say their family photos were used to generate child sexual abuse material through Grok, and to name Stability AI as a defendant over open models allegedly trained on such content. The complaint says xAI obstructed a law-enforcement investigation — by early 2026, the National Center for Missing and Exploited Children found 90 percent of xAI's tip-line reports were unusable because the company withheld the user information needed to identify perpetrators.

Prime Intellect raised a $130 million Series A at a $1 billion valuation to help companies train and run their own AI agents without handing proprietary data to OpenAI or Anthropic. Radical Ventures led the round, with the venture arms of Nvidia, Intel, and Dell joining. Founded in 2024, the startup sells a modular "full stack" — compute, a reinforcement-learning framework, and evaluation tools — that customers like Ramp and Zapier can adopt piece by piece rather than buy all at once.