Moonshot AI has released Kimi K3, an open-source frontier model featuring 2.8 trillion parameters, a 1 million token context window, and native multimodal capabilities, achieving efficiency breakthroughs in attention mechanisms. Official benchmarks show it outperforms Claude Opus 4.8 in some tests and competes with Fable 5 and GPT-5.6 Sol. Kimi K3 climbed 17 positions on Arena.ai's frontend coding leaderboard to claim first place, dominating 6 out of 7 design categories. Priced at $3 per million input tokens and $15 for output, it is the most expensive Chinese AI model, comparable to the Sonnet series. Emad Mostaque estimates its training cost at only $15–25 million, approximately 1e25 FLOPs, challenging the assumption that frontier models require massive computational resources. Real-world testing shows strong frontend generation quality but significantly slow speed—generating one page takes about an hour.

According to TLDR, Anthropic is preparing for an IPO and has scheduled investor meetings led by Goldman Sachs, Morgan Stanley, and JPMorgan Chase. The company recently completed a funding round valuing it at $965 billion, exceeding OpenAI's valuation. This marks a significant step toward public markets for leading AI firms. Additionally, Anthropic has launched a response strategy against OpenAI's 「reverse federalism」 approach, aiming to influence U.S. AI governance by pushing for stricter AI regulations state by state.

Payment giant Stripe and private equity firm Advent International have jointly proposed acquiring PayPal at $60.50 per share, implying a valuation of approximately $53 billion. They have secured around $50 billion in committed bank financing, and the two parties plan to split PayPal’s equity equally. PayPal’s current market cap is about $42 billion, down from previous highs above $300 per share. The acquisition remains in early stages and aims to achieve scale expansion while leveraging AI to improve efficiency across payments, Venmo, cryptocurrency, and cross-border transactions.

Moonshot AI revealed that Kimi K3 autonomously designed a two-stage kernel algorithm during a 15-hour self-evolution session, reducing forward and backward propagation time from 283.6 milliseconds to 114.4 milliseconds—a roughly 2.5x speed improvement. Additionally, its Vision-in-the-Loop capability integrates 3D reasoning, coding, and vision, enabling iterative transformation of concepts, images, and videos into playable interactive experiences via code-screenshot loops. These advances are seen as key progress in model self-improvement and multimodal agent development.

Cognition has announced the Devin Founder Program, offering startups $65,000 in usage credits for its AI software engineer across Cloud, Desktop, and CLI platforms. Selected startups will also gain access to exclusive events, merchandise, and white-glove support. The initiative aims to accelerate adoption of AI coding agents within early-stage entrepreneurial teams.

Thibault Sottiaux of OpenAI disclosed a critical bug in GPT-5.6 within Codex that can accidentally delete user files. The issue most commonly occurs when three conditions coincide: full permission mode is enabled without sandbox protection (including disabled auto-review), the model attempts to override the $HOME environment variable to define a temporary directory, and mistakenly treats the actual $HOME as a temporary folder to clean and delete. This configuration may result in deletion of the user's home directory and data loss. The official team advises avoiding such high-risk settings.

Google has officially renamed its AI research tool NotebookLM to Gemini Notebook, integrating it more deeply into the Gemini ecosystem—including the Gemini app and Google Search. Meanwhile, the hosted agents feature in the Gemini API now includes new cost-control and scheduling capabilities such as a free tier, budget guardrails, and time-based execution triggers. Gemini has also launched the Nano Banana Avatar feature in select countries, allowing users to generate custom avatars without re-uploading selfies each time.

Prediction market platform Kalshi has introduced AI compute futures contracts based on future GPU rental prices, enabling AI companies to hedge against fluctuations in computing costs, with plans to build a full derivatives market. This reflects the financialization of AI compute as a tradable asset. Concurrently, AI private banking startup Flex raised $70 million in funding, reaching a $1.2 billion valuation; Visa launched an AI financial assistant for banking apps, and research shows over 66% of Americans using generative AI rely on it for financial advice.