Grok 4.5 is SpaceXAI's first model release since going public: a multimodal workhorse with a 500K-token context window built for coding, agentic tasks, and office work. Benchmarks show it trading blows with Opus 4.8 on coding suites while producing roughly 4× fewer output tokens per task, making it the cheapest competitive option among frontier-adjacent developer models today.
Daily briefing · Auto-updated
AI News
Top stories from across the AI world, curated daily.
Unlike previous voice modes that ran in half-duplex (speak, then listen), GPT-Live-1 processes input and generates output concurrently, enabling natural interruptions, backchannel cues, and conversational turn-taking without forced pauses. For complex queries it offloads to GPT-5.5 in the background; developer API access is pending a waitlist.
Effective July 8, Fable 5 no longer draws from Pro, Max, Team, or Enterprise plan allowances; users who want continued access must opt in to pay-per-use credits at exactly double the Opus 4.8 rate. After immediate user backlash, Anthropic extended subscription-included access until July 12 while it communicates the change more clearly.
OpenRouter data shows DeepSeek and Z.ai's GLM consistently commanding nearly half of enterprise token volume at US companies, with prices 60–90% below US frontier alternatives; a coding session costing $10 on Claude runs under $0.50 on DeepSeek. At least one AI startup, Lindy, has migrated 100% of its traffic away from Anthropic to save millions, and Z.ai's GLM 5.2 saw 80× customer growth in its first week after launch.
At a July 2 internal all-hands, Zuckerberg acknowledged the AI agent development program has stalled despite Meta committing $125–145B in 2026 capex and laying off 10% of its workforce in May to clear headcount for AI-focused teams. The admission is notable given the scale of the bet and highlights how unpredictable frontier AI timelines remain even for organizations with near-unlimited resources.
Led by General Catalyst with Jeff Bezos and Eric Schmidt participating, the round backs an approach that uses billions of gameplay clips from Medal.com instead of expensive real-world robot demonstrations to build physical AI foundation models. In a live demo, the same model playing a video game for 100 hours was fine-tuned on 8 minutes of real-world footage to control a quadruped — suggesting game-derived spatial intuition may transfer to physical systems more readily than expected.
Sam Altman has pitched the idea directly to Trump, Lutnick, and Bessent: all leading US AI developers — including Anthropic, Google, and Meta — would cede 5% stakes to a fund modeled on the Alaska Permanent Fund, which pays dividends to citizens. The proposal is in early stages, would likely require an act of Congress, and appears designed as a political cover mechanism as the labs navigate an increasingly interventionist regulatory environment.