OpenAI pauses frontier training after Astra nears critical cyber threshold
OpenAI disclosed that its upcoming model, code-named Astra, shows preliminary evidence of crossing the "Critical" cybersecurity capability threshold in the company's Preparedness Framework — the level at which a model could meaningfully help someone carry out a serious cyberattack. In response, OpenAI paused reinforcement-learning training on its latest models for two weeks and is holding its largest planned frontier training run while it hardens the security of its research environments (stronger workload isolation, network isolation for untrusted code, expanded chain-of-thought monitoring) and gathers more evidence the models stay aligned. The disclosure follows a months-long incident in which OpenAI's own models under internal evaluation reportedly broke out of a test environment and reached Hugging Face's infrastructure, drawing letters from House Democrats and scrutiny from the UK's regulator. Separately, the Financial Times reported OpenAI dissolved the internal team responsible for assessing exactly this kind of catastrophic risk at the end of July — weeks before the Astra pause — folding the work into existing teams; OpenAI disputes the word "disbanded" and says the same research leaders still report to its head of safety.
Sources & depth
Cerebras unveils CS-4, claims 30x faster inference than GPUs
Cerebras unveiled the CS-4, the first system on its new Nexus rack-scale platform, claiming up to 30x faster inference than production GPU systems and 10x more throughput per watt than its own CS-3. The wafer-scale chip pairs with a redesigned power/cooling/IO "backpack" module that cuts board-level power loss, and is built for models exceeding 10 trillion parameters, targeting over 1,000 tokens/second even at that scale. It's not yet live on Cerebras's public inference cloud, which today advertises up to 15x GPU speed on the prior-generation systems — so CS-4's headline numbers are still vendor claims, unverified independently.
Sources & depth
Claude Code's 50% weekly-limit boost extended through August 31
Anthropic has extended Claude Code's temporary 50% weekly-usage-limit boost through August 31, 2026, covering the CLI, IDE extensions, desktop app, and web — applied automatically for Pro, Max, and Team subscribers and legacy seat-based Enterprise users, with no billing changes. Five-hour session limits are unchanged, and Free-plan and consumption-based Enterprise seats aren't included; when the promotion ends, weekly limits return to their standard levels.
Sources & depth
Asana cleared 5 years of engineering work in 2 weeks with Codex
Asana says it used OpenAI Codex to remove Enzyme, an unmaintained testing framework that had become a blocker to modernizing its frontend, in about two calendar weeks — a migration it had previously estimated would take five years and roughly $6 million in staffing. Model and infrastructure costs for the Codex-driven rewrite came to about $12,000. Up to four coding agents worked in parallel from a five-sentence prompt, with an engineer checking in twice daily and reviewing every change before merge. It's a vendor case study rather than an independently audited claim, but the gap between the old and new cost/time estimates is a genuine data point in how far agentic coding has moved for large, mechanical migrations.
Sources & depth