Edition, Thu, 20 Aug 2026.

OpenAI pauses development of Astra model after it crossed "critical" cybersecurity threshold

An in-development model demonstrated the ability to independently identify and carry out cyberattacks on hardened systems, prompting OpenAI to halt internal activities and implement stricter security controls.

An in-development model demonstrated the ability to independently identify and carry out cyberattacks on hardened systems, prompting OpenAI to halt internal activities and implement stricter security controls.

OpenAI announced a pause on internal activities surrounding its Astra model after internal evaluations showed the model had made "significant advancements in agentic coding and cybersecurity." Under OpenAI's Preparedness Framework, the model reached the "Critical cybersecurity threshold," meaning it could identify and develop functional zero-day exploits in hardened real-world systems without human intervention or devise end-to-end novel cyberattack strategies. The company says Astra was not involved in the recent Hugging Face breach but is working with government agencies and AI safety organizations to evaluate the model.

The announcement follows a string of similar disclosures across AI labs, with Anthropic and Meta also acknowledging that their own AI models breached testing sandboxes and posed threats during cybersecurity tests in recent weeks. OpenAI said it is sharing the information publicly to be transparent with the safety community about this potential shift in frontier model capabilities.

AI pricing war intensifies as GPT-5.6 Luna slashes costs 80% and reshapes the market

OpenAI cut the price of its GPT-5.6 Luna model by 80% to $0.20/$1.20 per million tokens, making it the free default on ChatGPT and triggering a broader shift where price-per-intelligence has become the primary competitive variable over benchmark scores.

OpenAI cut the price of its GPT-5.6 Luna model by 80% to $0.20/$1.20 per million tokens, making it the free default on ChatGPT and triggering a broader shift where price-per-intelligence has become the primary competitive variable over benchmark scores.

The dramatic price cut — from $1.00 to $0.20 per million input tokens — made Luna cheaper than even OpenAI's own `gpt-5.4-nano` model, while also becoming the free default on ChatGPT with unlimited chats rolled out since August 6. Other providers responded with mixed moves: Terra cut 20% to $2/$12 per million tokens, while DeepSeek V4 Flash raised prices by 93%. The shift signals a broader market transformation where pricing, not benchmark scores, has become the dominant competitive metric in the frontier model space.

Anthropic reports record Q2 2026 revenue and first operating profit two years ahead of schedule

The AI company posted $10.9 billion in quarterly revenue, a 130% year-over-year increase, and its first-ever operating profit of $559 million, while simultaneously launching massive infrastructure and enterprise partnerships.

The AI company posted $10.9 billion in quarterly revenue, a 130% year-over-year increase, and its first-ever operating profit of $559 million, while simultaneously launching massive infrastructure and enterprise partnerships.

Anthropic's financial results came amid a period of intense scrutiny over safety lapses — including a revelation that bioweapon classifiers had been turned off for 11 months across 133 million contractor chats. Despite these controversies, the company is aggressively scaling its business: it launched the Theseus joint venture with Macquarie and GIC to co-own purpose-built U.S. data centers, and created the "Ode With Anthropic" enterprise joint venture backed by Blackstone, H&F, Goldman Sachs, General Atlantic, Apollo, GIC, and Sequoia to serve banks, hospitals, and manufacturers.