AI Updates on 2025-08-06

OpenAI releases gpt-oss-120b and gpt-oss-20b as their first open-weight models in five years, with the 120B model built for production-grade applications with high reasoning capabilities and the 20B model for lower latency needs @AndrewYNg
Qwen releases Qwen3-4B-Instruct-2507 and Qwen3-4B-Thinking-2507 with 256K context length, featuring boosted general skills and advanced reasoning capabilities @Alibaba_Qwen
Perplexity adds Claude Opus 4.1 Thinking to their Max subscription service @perplexity_ai
OpenAI announces a livestream event for Thursday 10AM PT, with speculation about GPT-5 release @OpenAI

OpenAI is in early-stage discussions about a stock sale ahead of a potential IPO that could value the company at about half a trillion dollars @AndrewCurran_
OpenAI provides ChatGPT access to the entire U.S. federal workforce for essentially no cost ($1 per year per agency) through partnership with Government Services Administration @gdb
Google offers free Gemini Pro plans for college students in select countries for one year, plus $1B in funding for education and research @sundarpichai
Anthropic reports flying past $5 billion in ARR, making it one of the fastest-growing businesses of all time with focus on B2B applications @collision
ARR per employee emerges as the new startup metric that VCs are asking for earlier in company lifecycles as a measure of capital efficiency @GergelyOrosz
AI coding tools raise the floor but not the ceiling of software development, making it easier to create mediocre software but not enabling great software by itself @GergelyOrosz

Google DeepMind publishes research on developing new ethical frameworks for AI agents as they begin taking action in the real world, emphasizing alignment with well-being and societal norms @GoogleDeepMind
Anthropic updates Claude's system prompt to address sycophancy issues, allowing it to be more critical of user theories and break character in roleplaying when appropriate @AmandaAskell
The system prompt changes also help Claude be more direct about mental health concerns and avoid agreeing its way into existential distress @AmandaAskell

Claude Code now automatically reviews code for security vulnerabilities and integrates with GitHub Actions for automatic reviews on every pull request @claudeai
Google's AI coding agent Jules exits beta and becomes generally available as an asynchronous coding agent that can check out repos and submit pull requests @simonw
Microsoft introduces Copilot Vision for Motorola users in moto ai, enabling visual assistance in 50+ languages for tasks like translating street signs @mustafasuleyman
Perplexity Finance charts are described as a piece of art that makes users unable to use other finance products @AravSrinivas
Google launches new Guided Learning mode in Gemini with visual aids, quizzes, and conversational explanations to help students understand and retain information @GeminiApp

OpenAI's gpt-oss-120b model required 2.1 million H100-hours to train, with estimated costs between $4.2M and $23.1M based on H100 pricing ranges @simonw
The new OpenAI open-weight models are considered to hold their own or even beat models from Chinese AI labs over recent months @simonw
Microsoft Research introduces VeriTrail, which can detect AI-generated content not supported by source text and trace content provenance back to sources @MSFTResearch
Microsoft pioneers a vision for self-adapting AI systems that can adapt to the dynamic nature of scientific discovery for deeper reasoning in complex scientific domains @MSFTResearch
PyTorch 2.8 releases with limited stable libtorch ABI for third-party C++/CUDA extensions and high-performance quantized LLM inference on Intel CPUs @PyTorch