AI News: June 7, 2026

1. Anthropic Holds Top Two Arena Spots as Meta’s Muse-Spark Debuts at Number Three

LMSYS Chatbot Arena. Anthropic occupies the top two positions on the Chatbot Arena text leaderboard, with claude-opus-4-6-thinking at 1503 and claude-opus-4-7-thinking at 1500, ahead of Google’s Gemini 3.1 Pro Preview at 1488 and OpenAI’s GPT-5.5 at 1482. Meta’s muse-spark enters at number three with a score of 1489, a notable new appearance for a Meta model near the top of the human-preference rankings. With the leading six models separated by roughly 20 points, the frontier is now tightly clustered on blind preference voting. Source

2. Meta Builds Hatch, Reportedly Its First Paid AI Product at Up to $200 a Month

Meta. According to a report, Meta is developing Hatch, a subscription AI agent that automates tasks such as tool creation and email handling, with pricing said to reach as high as $200 per month. It would mark Meta’s first paid AI offering, signaling a shift from free consumer AI features toward direct monetization. The reported pricing would place Hatch in the premium agent tier alongside competitors’ top-end plans. Source

3. Trump Administration in Talks to Take an Equity Stake in OpenAI

OpenAI. President Trump indicated that discussions are underway about the U.S. government taking an equity stake in OpenAI, with proposals involving a public fund to distribute benefits to citizens. Lawmakers have separately floated heavy taxation on AI company shares. Government ownership of a leading frontier lab would be an unprecedented structural development for the sector and could reshape how AI firms are regulated and financed. Source

4. xAI Reportedly Trained Coding Models on Claude Outputs Before Being Cut Off

xAI. A report says xAI used outputs from Anthropic’s Claude to train its coding models for months before its access was restricted. The piece also describes a shrinking pretraining team at xAI and compute being redirected elsewhere. The account raises fresh questions about model distillation practices and terms-of-service enforcement among frontier labs. Source

5. Sakana AI Bets on Recursively Self-Improving Models Over Raw Compute

Sakana AI. Sakana AI, co-founded by a co-author of the original Transformer paper, launched a research effort centered on recursively self-improving AI as an alternative to scaling raw compute. The thesis is that models capable of improving themselves could erode the capital advantage currently held by the largest frontier labs. It is a research direction to watch rather than a shipped product. Source

6. New Open-Source Voice Model Decides Every 0.4 Seconds Whether to Speak

Open source. A new open-source audio model performs real-time transcription and translation, continuously deciding in roughly 0.4-second intervals whether to speak or stay silent rather than waiting for a recording to finish. The code is released under the Apache 2.0 license. The streaming, turn-taking design targets low-latency voice agents that need to interleave listening and speaking. Source

7. Nadella Publicly Rejects an Internal Plan to Make Microsoft’s Agent Addictive

Microsoft. Microsoft CEO Satya Nadella publicly rejected an internal proposal to design the company’s Scout agent for user dependency, stating that AI should empower users rather than addict them. The episode surfaces internal tension over engagement-maximizing design in AI products. It feeds a wider debate about what responsible agent user experience should look like as assistants become more capable. Source

8. Sebastian Raschka Publishes a Curated 2026 LLM Research Reading List

Ahead of AI. Sebastian Raschka published a curated compilation of notable LLM research papers from the first half of 2026, organized by category including reasoning models, efficient inference, reinforcement learning, and agent systems. The list is a practitioner-oriented reading resource rather than new research. It serves as a survey for teams tracking the year’s architecture and training trends. Source