ENFR
8news

Tech • IA • Crypto

TodayMy briefingVideosTop articles 24hArchivesFavoritesMy topics

ChatGPT 5.5 hits 1000 tok/s, GPT Real-Time 2 voice leap

AIMonday, May 11, 2026· 10 videos

Briefing

0:00 / 0:00

ChatGPT 5.5 shifts to Cerebras

ChatGPT 5.5 introduces a new inference stack built on Cerebras processors, moving beyond NVIDIA H100/H200 GPUs. Generation speeds reportedly jump from about 65 tokens/sec to 1000 tokens/sec, enabling near real-time output. The system also adopts post-inference task overlap, allowing parallel execution of steps instead of linear processing. This architecture underpins more complex, continuous workflows rather than one-shot responses.

GPT Real-Time 2 goes voice-native

OpenAI launches GPT Real-Time 2, enabling simultaneous speaking, listening, and reasoning in live conversations. The model integrates GPT-5-class reasoning and expands context to 128,000 tokens, supporting longer, coherent dialogue. It can call tools mid-conversation, connecting to systems like CRMs, calendars, and databases. Deployment is API-first, signaling an initial focus on developers and enterprise use.

Claude Mythos hits 16-hour autonomy

Claude Mythos reportedly achieves a 16-hour task horizon, far exceeding prior evaluation limits. The METR benchmark struggles to measure it, with only 5 of 228 tasks long enough to test its ceiling. Capability growth has accelerated from seconds in 2021 to hours in 2024–2026, suggesting a super-exponential trend. Researchers warn this marks a shift toward autonomous digital workers capable of extended independent execution.

Agentic AI becomes dominant paradigm

Agentic systems are rapidly replacing prompt-response models across platforms like ChatGPT 5.5 and Claude 4.7. These systems operate in loops involving context gating, decision-making, and verification, rather than single outputs. Tasks are dynamically routed to tools and executed across multiple steps with persistent memory. The shift reframes AI as an orchestrator of workflows instead of a text generator.

Abacus AI builds design agents

Abacus AI unveils an agent-driven platform that executes full creative workflows from concept to final assets. Its system performs structured reasoning before rendering, generating outputs via Python-based design pipelines. In one case, it produces 30+ screens for a credit onboarding flow, including edge cases and recovery paths. The approach emphasizes continuity, turning rough sketches into production-ready interfaces.

China rejects AI-driven layoffs

Chinese courts are ruling that companies cannot justify layoffs solely by replacing workers with AI. Judges state AI adoption is not an “external shock” under labor law, blocking its use as a legal excuse for dismissals. Firms must offer reassignment or retraining instead of termination. The rulings aim to balance rapid AI expansion—now involving 4,500+ companies—with workforce protection.

OpenAI evolves into tech superpower

OpenAI, founded in 2015 as a nonprofit, is now valued near $500 billion with strong backing from Microsoft. Its shift toward closed systems and global infrastructure marks a departure from early open ideals. CEO Sam Altman has outlined ambitions where AI systems coordinate economic and societal functions. The company’s scale increasingly resembles state-like influence over digital ecosystems.

Claude Co-work enables persistent workflows

Claude Co-work introduces a workspace-centric model where AI operates directly on local files and folders. Persistent layers like Claude MD and Memory MD store instructions and long-term context across sessions. Projects are segmented into domains, each with its own memory and rules, preventing context bleed. This structure turns AI into an operational system for ongoing business processes.

Videos covered

Previous briefings · AI