OpenAI chief scientist Jakub Pachocki says chain-of-thought monitoring — watching models write out their reasoning to catch misalignment — is becoming unreliable, and no lab is ready to scale safely.
Perplexity's own research-agent benchmark now crowns GPT-6 Astra, which scores 13.5% higher than Claude Fable 5.1 — while costing 6.1% less per task.
Nvidia is shelling out $6 billion to license Poolside's Model Factory and scoop up 109 employees who worked on the Laguna model — while Poolside keeps its independence.
DeepSeek's first Agent product, Harness, is out on npm — a build-it-yourself agent toolkit that goes beyond coding to handle multi-agent workflows, skills, and web search.
RuntimeWire found that Kimi Work automatically attaches your five most recently updated agent sessions to every feedback submission — without showing them or asking permission.
Stripe is reportedly in talks to buy OpenRouter for nearly $10 billion, according to The Wall Street Journal — a massive jump from the AI aggregator's $1.3 billion valuation just two months ago.
DeepSeek open-sourced Harness and revealed the Cordis design paper, explaining how making every component — including the agent loop — a swappable plugin paves the way toward recursive self-improvement.
DeepSeek open-sourced Harness and revealed the Cordis design paper, explaining how making every component — including the agent loop — a swappable plugin paves the way toward recursive self-improvement.
Terminal-Bench 3.0 has a new leader: Claude Opus 5 Max + mini-SWE-agent tops the board at 43.5%, beating GPT-5.6 Sol Max with Codex. But the benchmark measures the model and agent together, not the model alone.
Terminal-Bench 3.0 has a new leader: Claude Opus 5 Max + mini-SWE-agent tops the board at 43.5%, beating GPT-5.6 Sol Max with Codex. But the benchmark measures the model and agent together, not the model alone.
