75 OpenAI employees just got $30M each

TechCrunch

Adaption has introduced AutoScientist, an automated fine-tuning tool that co-optimizes data and models to help AI systems train themselves more efficiently, potentially democratizing frontier AI development beyond major labs. The company claims it has doubled win rates and is offering free 30-day trials.

Microsoft's multi-agent MDASH system topped the CyberGym cybersecurity benchmark at 88.45%, beating Anthropic's Mythos and GPT-5.5 by using over 100 specialized AI agents to find real vulnerabilities including 16 new Windows flaws. The advancement demonstrates both AI's defensive security potential and raises concerns about its offensive hacking applications.

Kilo AI

DeepSeek's new V4 models establish distinct positions: V4 Pro outperforms Kimi K2.6 at lower cost (77 vs 68 score), while V4 Flash creates an unprecedented ultra-budget tier at $0.02 per complex backend generation—100x cheaper than Claude Opus 4.7 despite lower quality, potentially changing economics for iterative workflows.

Anthropic

The article link is broken (404 error), but the title suggests Anthropic may be announcing an 'Agent View' feature for Claude Code, potentially enhancing Claude's coding capabilities or interface.

OpenAI enabled 600+ employees to cash out $6.6 billion in shares through a tender offer, with 75 receiving the maximum $30 million payout. This marks one of Silicon Valley's largest pre-IPO wealth creation events, as the company stays private while providing liquidity to employees.