Sep 4 · AI Daily

According to Artificial Analysis data, Grok 4.6 (High) beats OpenAI GPT-6 Astra (Max) on several benchmarks: Agentic Index 59 to 51, τ3-Banking 51% to 41%, Non-hallucination 66% to 49%, and GDPval-AA v2 1,730 to 1,629. The author notes that Grok 4.7 is coming soon.

Anthropic's Claude Devs team announced the addition of the ant apply command to the ant CLI. Users can declare Claude Managed Agent environments, agents, skills, memory storage, and deployments as files in a repository, then use ant apply to keep API resources in sync with those files. Documentation is available at https://platform.claude.com/docs/en/cli-sdks-libraries/cli/apply.

The author announced that starting today, paid ChatGPT plan users will receive one banked reset for each day they are unable to access Astra. The first round of compensation will be issued in about 3 hours, and the team is working hard to restore access as soon as possible. Users who haven't registered an account yet can still sign up.

The Anthropic team is exploring new ways to extend and customize Claude Code through Function Hooks, sharing a demo video to show how they work. The feature is not yet released, and the team is collecting feedback on a GitHub issue.

Ant inclusionAI released Choruz on GitHub, a local-first collaboration app where humans and AI agents work together in Slack-like spaces. Each agent runs a real CLI in an independent workspace and hands off work via messages, threads, tasks, and files.

Gemini Live now works with Google Workspace connectors, including Gmail, Keep, and Docs. The author notes this is a long-awaited feature change.

ChatGPT, Claude, Codex, and Grok suffered service disruptions in the same time window, and both OpenAI and Anthropic confirmed elevated error rates.

Emad Mostaque shared that Astra was pre-trained using 100k GPUs at Stargate Texas over roughly two months, making it the first $1-billion-scale training run.

The BestBlogs 09-04 morning report rounds up ten items covering the GPT-6 Astra security overview, a startup interview with Paul Graham, Zeng Ming's perspective on the AI industry, an open-source harness comparison between Codex and DeepSeek, Meta's team reduction plans and organizational second brain, an NVIDIA speculative decoding guide, Qwen E-Commerce Bench, Google WeatherNext 3, and Meituan Smart Broadcast.

Author Digital Life Kazuk released a comprehensive analysis of GPT-6 Astra, titled with the theme "Welcome to the AGI Era." Current materials only provide summary-level information, and the main text describes it as "a gradual journey" without providing verifiable specific details.

Cognition announced that GPT-6 Astra is coming to Devin, scoring within 0.4 points of Fable 5 on FrontierCode 1.1 at a 64% lower cost, while setting a SOTA on its internal test benchmark by generating more thorough tests, clearer reports, and better video evidence.

OpenAI announced the Daybreak for Frontline Defenders initiative, committing $1 billion (about 6.736 billion yuan) to provide resource-constrained cybersecurity defense teams with Daybreak access, training, technical support, and collaboration services, prioritizing critical services such as water supply, electricity, local government, and finance.

The Los Angeles Unified School District (LAUSD) has blocked generative AI for all students on district devices; Google Search remains available, but AI Mode is disabled, and AI Overviews may still appear because Google has not yet provided administrators with a complete off switch. Teacher use of AI is unrestricted.

SemiAnalysis suggests that OpenAI's launch of an ASIC chip project, investing hundreds of millions in R&D and showcasing Jalapeno at Hot Chips, is a good way to pressure NVIDIA to secure billions of dollars in financing and guarantees. If Jalapeno performs well at scale, OpenAI's revenue per MW will rise; even if the chip fails, OpenAI can still secure substantial financing and guarantees from NVIDIA.

Ryan Greenblatt states that GPT-6 Astra has made a huge leap in opaque reasoning, capable of mentally solving difficult contest math problems without showing its work, whereas previous AI could only solve basic word problems. He warns that benchmark results may be contaminated and mentions that UK AISI found Astra's monitorability is notably worse. He speculates the jump comes from architectural changes that increase serial depth, and warns that if future model generations continue this trend, chain-of-thought will no longer be an effective oversight tool.

Citing GPT-6 Astra sitting on the Pareto frontier of the Artificial Analysis intelligence index due to extreme token efficiency, with a per-task cost lower than Gemini 3.8 Flash despite being 13 times cheaper than that comparison, Emad Mostaque predicts frontier models will one-shot everything 100 times faster and cheaper than today by next year. He believes that since monitoring reasoning traces won't be necessary at that point—representing instinct-like intelligence—the content embedded during pre-training will matter even more.

The BestBlogs 09-04 morning report details three items: OpenAI's release of the GPT-6 Astra security overview, calling it the first model in its preparedness framework to cross the critical cybersecurity capability threshold, and detailing changes in jailbreak resistance, mis-alignment monitoring, and chain-of-thought monitoring; Paul Graham discussing how AI changes information access and startup cost structures in a YC interview, with compute becoming a new line item; and Zeng Ming discussing AI industry and organizational evolution through infrastructure, application exploration, and native application frameworks.

Gary Marcus published a review of GPT-6 Astra, stating that multiple reports show it is real progress and that OpenAI's product explicitly creates and manipulates symbolic world models, validating his arguments from the past decade.

Ethan Mollick suggests that the mental model for using Fable and Astra-tier models is delegating to a capable external team rather than an intern. You need to be able to articulate what you want, give the AI a clear scope of autonomy, specify what to test and when to ask for help, and describe what a good outcome looks like.

Zeineb Farhat, a researcher at the Paris ISC Business School, published a small study in Applied Cognitive Psychology interviewing 45 users aged 18 to 25 who use ChatGPT daily. The study found that over-reliance may be linked to decreased independent thinking, increased self-doubt, and social isolation, connecting it to cognitive offloading.