OpenAI made GPT Transcribe and GPT Live Transcribe generally available on July 28, 2026. The new default path separates completed-file transcription from low-latency live audio and adds prompt, keyword, and multi-language hints.
AI
RSS FeedOpenAI said frontier model acceleration may eventually need deliberate pacing tools. The linked petition lists 1,224 employees from frontier AI companies and asks the U.S. government to support an international effort.
OpenAI made Codex Security CLI and its TypeScript SDK open source. The tweet drew more than 640,000 views, and the tool covers repository scans, finding history, fix checks, and CI security gates.
The HN debate centered less on Anthropic’s denial of a ban and more on whether mandatory safety testing could become a de facto gate.
The new security race is less about one giant model and more about routing work to the cheapest capable model. Microsoft says MAI-Cyber-1-Flash inside MDASH reaches 95.95% on CyberGym while cutting cost by about 50%.
OpenAI expanded GPT-Live-powered ChatGPT Voice to Edu, Business, and Enterprise plans globally. The help center lists 5 credits per minute for Business Voice in Chat and about 6 credits per minute for Voice in Work and Codex.
Safe Superintelligence will expand its compute by an order of magnitude through NVIDIA investment and access to Vera Rubin systems. The deal turns a productless frontier lab into a strategic customer for upcoming AI infrastructure.
NVIDIA launched the Open Secure AI Alliance with Microsoft, Cloudflare, Hugging Face, Palantir, and other partners. The bet is that AI-agent defense needs open models, harnesses, logs, and evaluation tools that defenders can inspect and run themselves.
OpenAI analyzed more than 800,000 U.S. ChatGPT work messages and found that 43.5% of occupation-specific use involved tasks from another job. The data points to AI changing who does the work before job titles catch up.
AI safety scrutiny is shifting from abstract risk debates to incident review and public technical reporting. OpenAI said on July 25 that the Hugging Face-related incident is under review with external advisers and committee oversight, drawing more than 303,000 views.
The discussion moved quickly from sample quality to the larger claim: one multimodal backbone for visual generation and action prediction.
MCP moves to a stateless core on July 28, 2026, and GitHub MCP Server already supports the latest spec. The practical changes are fewer session dependencies, less payload inspection, and official conformance tests for agent infrastructure.