OpenAI Fires Safety Researchers; GPT-6.1 Sol Ultrafast Debuts
The dismissals raise questions about third-party safety audits even as OpenAI ships a faster GPT-6.1 tier and Anthropic cuts cache-read prices.
AI & ML
Frontier models, research results, agents, and developer tooling.
Gears of War: E-Day Streams on GeForce NOW With DLSS and RTX
NVIDIA's cloud gaming service adds the new Gears prequel with RTX 5080-class performance and AI-driven DLSS, plus a direct Fire TV membership purchase option.
JetBrains Mellum2.1: 12B MoE Coding Model Hits 47.0 on SWE-bench
JetBrains' open-weight Mellum2.1 uses reinforcement learning on real repositories to jump from 2.0 to 47.0 on SWE-bench Verified.
NVIDIA's PivotOPD Helps AI Agents Recover From Early Mistakes
A new on-policy distillation method trains multi-turn LLM agents to avoid pivotal mistakes and recover from them, outperforming 13 baselines on three benchmarks.
Architect's Liquid Inference Runs Live Auctions for LLM Requests
A new marketplace lets LLM providers bid in real time to serve each inference request, aiming to cut costs.
Nvidia expands Halos safety platform from self-driving cars to robots
Nvidia's full-stack Halos safety system aims to reduce risks in physical AI, from robotaxis to humanoid robots.
Anthropic's Claude Haiku 5.5 Offers 1M Context at $0.10 per Million Tokens
The new small model undercuts typical pricing while retaining a long context window and a strong OSWorld score.
Unsloth Studio Re-Checks Trusted Model Repos Before Execution
Unsloth Studio binds approval to code fingerprints and blocks flagged weights, closing the gap between a trusted repo and the code that actually runs.
Perplexity Releases pplx-embed-v2-late Embedding Models
Perplexity's new open-weight embedding models span edge to server, with a top MADQA score of 92.4%.
Late-Interaction Embeddings Arrive in Sentence Transformers
Hugging Face's new support for multi-vector models gives retrieval systems a more granular matching method than traditional single-vector embeddings.
Asana completes 5 years of testing work in 2 weeks with OpenAI Codex
A single source reports that Asana used OpenAI's Codex to finish a legacy testing overhaul in two weeks for about $12,000, work previously estimated at five years.
Tutorial Offers Three Ways to Handle Small LLM Context Windows
A practical guide demonstrates how to work around limited context in language models using code examples.
Agent Memory Needs: New Guidance from IBM and Hugging Face
A Hugging Face post from IBM Research offers a practical framework for sizing agent memory, challenging one-size-fits-all assumptions.
OpenAI Adds Safeguards to Pace Frontier Model Development
OpenAI is tightening monitoring, alignment, and security measures to control the release pace of its most capable AI models.
Google Search adds study tools for practice and quizzes
New features aim to turn Search into a study aid for students preparing for tests.
Replit Free Mode Uses OpenAI's GPT-5.6 Luna to Remove Token Barriers
Replit's new Free Mode, powered by OpenAI's GPT-5.6 Luna, lets anyone build working software without tracking token costs.
OpenAI Extends Zero Data Retention to Frontier Models
OpenAI reaffirms Zero Data Retention for eligible API customers and previews Private Safety Processing for advanced AI safety without compromising data privacy.
OpenAI pushes AI oversight, literacy, and policy research
Three new OpenAI initiatives span national-security oversight, student AI literacy, and independent policy research.