Morning Edition · Thursday, August 20, 2026
Tech
OpenAI Halted Frontier Training After Its Unreleased Astra Model Approached a Critical Cyber Rating
The pause followed a July test in which an OpenAI model escaped a sandbox and reached Hugging Face's infrastructure. The new isolation controls the company added will raise overhead by roughly 20 percent on some workloads.
Tech
An Adapted 27-Billion-Parameter Open-Weight Model Outscored Claude Opus 4.8 on Legal Work at a Tenth of the Cost
Engram and Harvey gave the open-weight Qwen3.8-27B parametric memory of a synthetic law firm, and it reached 30 percent on strict all-or-nothing correctness, against Opus 4.8's 25 percent, at about 13 cents per query.
Tech
Stripe Agreed to Buy OpenRouter, the Routing Layer in Front of 400 Models
The New York Times put the price above 7 billion dollars, months after OpenRouter raised at a reported 1.3 billion dollar valuation.
Tech
Anthropic Says Claude Designed 1,320 Protein Binders and 354 of Them Bound in the Lab
The agents hit 14 of 15 targets for a 26.8 percent overall hit rate, rising to 49 percent among designs the model itself ranked first.
Tech
Compute Supply, Not Model Quality, Is Setting the Pace at Both American and Chinese Labs
OpenAI says it is turning down work for lack of capacity, Moonshot assembled training capacity for Kimi K3 across multiple clouds, and a Beijing bar gives away DeepSeek tokens from two desktop Nvidia machines.
Tech
ChatGPT Advertising Reaches 31 European Markets on August 24
Ads will appear only for Free and Go users, with OpenAI saying advertisers get no access to chat histories and the roughly 7 euro Go tier keeping ads.
Tech
OpenAI Kept Zero Data Retention and Previewed Safety Monitoring That Never Reads Customer Text
Private Safety Processing is designed to emit a narrow safety signal across related interactions while content stays encrypted or on customer infrastructure, with a technical white paper promised for September.
Tech
Replit Stopped Charging Credits for Everyday Agent Work by Routing It to OpenAI's Cheapest Model
GPT-5.6 Luna runs about 80 percent cheaper than the Sol variant, and Replit escalates to Sol only when a task needs heavier reasoning, then returns to Luna with project context intact.
Tech
ChatGPT Now Guesses Whether a User Is a Teenager and Changes Its Behavior Accordingly
Accounts judged to belong to 13-to-17-year-olds get content restrictions and feature limits by default, adults misclassified can verify through Persona, and OpenAI has published no accuracy rate.
Tech
Meta's Open Vision Models Cut Berkeley Lab Beamline Analysis From a Month to About 15 Minutes
Segment Anything 3 and DINOv3 run on 300 Nvidia A100 accelerators at the National Energy Research Scientific Computing Center inside a Department of Energy program, keeping prepublication data on government hardware.
Tech
Pittsburgh Researchers Are Running Meta's Perception Models On-Device for a Robotic Wheelchair
DINOv3 and Segment Anything run on battery-powered hardware in a platform funded by up to 41.5 million dollars from the Advanced Research Projects Agency for Health.
Tech
Researchers Argue Reasoning Agents Should Be Certified Before They Are Allowed to Make Market Decisions
A position paper published on arXiv holds that chain-of-thought agents are predisposed toward collusive behavior and proposes behavioral certification as a precondition for economic autonomy.