AI Daily Newsletter: Grok 4.5 Officially Launches and Lovable Valued at 13 Billion USD

Grok 4.5 Officially Released by xAI and Cursor

xAI in collaboration with Cursor has released Grok 4.5 — a mixture-of-experts model trained on trillions of tokens of data from Cursor and SpaceX. This is the most intelligent model to date from xAI, capable of handling complex programming tasks, data science, finance, and law. On Hacker News, this post reached 401 points. With a price of $2/$6 for input/output, it is considered to be four times more cost-effective than Claude 3 Opus.

Source: Cursor Blog | xAI | HN

ChatGPT GPT-Live: Real-Time Conversation like a Human

OpenAI has launched GPT-Live, a new generation of voice models with a full-duplex architecture — capable of listening and speaking simultaneously. This model makes decisions multiple times per second: when to speak, when to listen, when to pause, or call a tool. GPT-Live-1 is preferred over 75.7% compared to the old Advanced Voice Mode. It is now available for Plus and Pro users.

Source: The Decoder

German Court: Google Responsible for AI Overviews

A historic ruling from the Munich Regional Court: Google has been determined to be directly responsible for the content of its AI Overviews. Google’s AI incorrectly linked two publishers with scam campaigns, despite no linked source mentioning this. The court believes that AI Overviews operate differently from regular search — Google rewrites and evaluates results using its “own language”.

Source: The Decoder

Mistral Launches Robostral Navigate: 8B AI Navigation Model

Mistral introduces Robostral Navigate, an 8B parameter model that achieves 76.6% on the R2R-CE benchmark with just a single RGB camera — no depth sensor, LiDAR, or multiple cameras needed. This model is specialized for mobile robots, aiming to build an “original” AI navigation system for robots.

Source: Mistral AI | HN (381 points)

pxpipe and SWE-1.7: Optimizing Token Costs and Benchmarks

Two other notable news items: (1) pxpipe — an open-source tool that hides text in PNG images to reduce token costs for Claude Code and Fable 5 by up to 70%, reducing session costs from $42 to $6. (2) SWE-1.7 from Cognition Labs achieves performance comparable to GPT 5.5 and Claude Opus, reaching 239 points on HN.

Source: The Decoder (pxpipe) | Cognition (SWE-1.7)

👔 Anthropic Turns Fable 5 into a “Manager” — Sonnet 5 Does the Rest

Anthropic has introduced a rather smart (and somewhat humorous) strategy to solve Fable 5’s cost problem: instead of running this token-hungry monster on every task, turn it into a manager — planning, dividing tasks, and then assigning them to Sonnet 5. This “Advisor” pattern has been officially recommended by Anthropic, resulting in maintained benchmarks but significantly reduced bills. Essentially, it’s similar to how AI startups often operate: using a strong model for planning and a smaller one for execution. This time, Anthropic admits its model is too expensive, so it delegates — a clever way to “cheat”.

Source: The Decoder

🕵️ Google’s Deepfake Detection System Exposes Fake McConnell Photo

A photo circulating online earlier this week showed Senator Mitch McConnell in a hospital bed with tubes all over him, looking very dire — it turned out to be entirely AI-generated. Google used its deepfake detection system to confirm this as fake and helped media outlets debunk it in time. This reminds us that as AI-generated content becomes easier, AI detection of fake content must also develop. This case raises questions about the responsibility of platforms in preventing misinformation during election seasons — no wonder Google is heavily investing in this area.

Source: TechCrunch

💰 Lovable to be Valued at $13.2 Billion — AI Coding Platform “Hottest” in the Market

Lovable, an AI platform that helps developers build applications quickly, is in a $300 million funding round led by Menlo Ventures — which could push its valuation up to $13.2 billion, doubling its previous valuation. In the context of a highly competitive AI coding tools market (Cursor, Lovable, and GitHub Copilot are all competing fiercely), this valuation shows that investors are betting big on the future of “AI writing code for developers”. It’s understandable — if a tool can help devs increase productivity 3-5 times, then $13.2 billion might not be too expensive after all.

Source: TechCrunch

This is the updated news for the afternoon of 09/07/2026. If there’s anything hot, I’ll update it further.