ENFR
8news

Tech • IA • Crypto

TodayMy briefingVideosTop articles 24hArchivesFavoritesMy topics

GPT Real-Time 2, Colossus 1 deal, Claude Mythos shock

AITuesday, May 12, 2026· 13 videos

Briefing

0:00 / 0:00

GPT Real-Time 2 goes live

OpenAI has introduced GPT Real-Time 2, a voice-first model capable of simultaneous reasoning and conversation. It integrates GPT-5-class intelligence with low-latency speech, enabling fluid, human-like dialogue. The system supports 128,000-token context and real-time tool use across external systems. Available via API, it signals a shift toward conversational interfaces as primary computing layers.

Anthropic secures Colossus 1 compute

Anthropic has gained access to SpaceX’s Colossus 1 supercomputer, reportedly featuring 220,000 GPUs. The deal dramatically expands compute capacity for Claude models, improving latency and scale. Deployment is expected within weeks, unusually fast for infrastructure of this magnitude. The move deepens Anthropic’s multi-cloud strategy across Amazon, Google, Microsoft, and SpaceX.

Claude Mythos breaks evaluation limits

Claude Mythos has reached a 16-hour task autonomy, exceeding current benchmark capabilities. The METR evaluation system lacks sufficient long-duration tasks, creating what researchers call an “evaluation crisis.” Capability growth has accelerated from seconds in 2021 to hours in 2026, suggesting super-exponential trends. The model blurs the line between tools and independent digital workers.

Codex expands to full computer control

OpenAI Codex now operates entire computer interfaces, interacting with apps via graphical controls. It can click, type, and navigate across tools like Spotify or messaging apps without APIs. The system supports parallel multitasking, running workflows simultaneously in the background. This marks a shift from coding assistant to general-purpose autonomous operator.

Codex 5.5 Chrome automates browsing

The Codex 5.5 Chrome extension enables direct browser automation on macOS and Windows. Users can extract data, navigate sites, and execute workflows like real estate research in real time. Reusable “Skills” allow multi-step task automation without manual input. The browser becomes an execution layer for AI-driven operations.

China courts reject AI layoffs

Chinese courts are ruling that companies cannot justify layoffs solely by adopting AI systems. Judges found automation is a strategic choice, not an “external shock” under labor law. Employers must offer reassignment or retraining instead of termination. The decisions balance rapid AI growth—across 4,500+ companies—with worker protections.

OpenAI adds Trusted Contact safety

OpenAI is rolling out Trusted Contact in the United States to address self-harm risks. The system monitors conversations for distress signals and can alert a designated contact. Escalation may include AI intervention and human review before notification. Alerts exclude message content, aiming to preserve privacy while enabling support.

Kiro 3.1 pushes local multi-model AI

Kiro 3.1 introduces a free, browser-based platform supporting local AI and multiple providers. It integrates OpenRouter, enabling access to models from OpenAI, Anthropic, Google, Mistral, and Perplexity. Users can switch models mid-task and combine outputs in one workflow. Local execution via Ollama and LM Studio emphasizes privacy and cost control.

Videos covered

Previous briefings · AI