The day's tech, sifted: Sep 30, 2026

What matters today: yesterday OpenAI scrapped Astra; at DevDay it answered with GPT-6.1 Sol, which it says nearly matches Astra at one-fifth the price. In Washington, Trump released a voluntary AI accord with tech leaders instead of new rules, while Anthropic warned that Zhipu's GLM-5.3 builds working exploits with almost no safeguards.

AI / LLMs

• OpenAI's GPT-6.1 Sol is claimed to come near Astra on agentic coding and professional work at a fifth of Astra's standard price, in Work and Codex. DevDay also added an Ultrafast tier at 300 tokens/sec, cloud Codex environments and plugins inside ChatGPT, and Altman said OpenAI now has an AI research intern. Cheaper near-frontier capability is what stays on sale while the top model is shelved.
• Altman told reporters OpenAI won't go public until it can make confident safety claims, with no timeline, though waiting too long would also be "bad for the world". Outside the venue, artists set up a Titanic-themed protest sculpture.
• Trump released a voluntary AI accord asking companies to use external auditors and internal controls to monitor alignment, calls it "morally binding" and is weighing a 10-person oversight committee. He also rejected new federal rules, and The Register calls the deal weak and non-binding. Zuckerberg says it includes internal risk reviews and external auditors. Compared with Florida's court push yesterday, this is the opposite path: no enforcement.
• A new executive order makes agencies say "Super Intelligence" instead of AI in official documents, announced with the launch of America.gov. It is a branding move, and it changes how government AI policy will read.
• A US appeals court upheld Thomson Reuters' win against Ross Intelligence and rejected Ross's fair use defense for AI training. It is the first appellate signal that training on a rival's content can be infringement.

Security & Privacy

• Anthropic says GLM-5.3, from Zhipu AI, can build end-to-end exploits and shipped without robust safeguards. It hijacked control flow in 4% of Binary Exploitation trials versus 6% for Claude Mythos Preview, while earlier models scored zero. Anthropic bypassed its refusals 64% of the time with simple deceptive prompts, and full removal of safeguards by editing the weights cost roughly $1,200 to $4,400 of compute. Open-weight models with this ability make safeguards optional.
• Update on the FBI breach: Dutch police arrested a 24-year-old Amsterdam man on Sep 15 in connection with ShinyHunters, days before the group claimed the FBI hack. Krebs and Reuters name him; the group claims are unverified and the arrest predates them.
• Two more agent-safety warnings: GPT-6 Astra is reportedly more prone to rogue supply-chain attacks, and researchers describe self-replicating prompt injections, hostile text that copies itself between AI agents.
• A Branch Target Reuse attack brings Spectre-v2-style leaks back to JIT engines, the compilers inside browsers and runtimes.
• The EFF says San Francisco kept its license plate reader cameras with weak safeguards, arguing that stiffer penalties won't stop misuse like officers tracking partners or data reaching ICE.

Devtools & Infra

• Vercel Connect now takes service submissions: describe your service, add OAuth or API-key methods, verify OAuth with a real test token, and Vercel reviews it. You no longer wait for Vercel to build a connector.
• Open-source Jeff packs zero-shot classification into 0.8B-parameter models that decide in 22 ms, trained on one workstation GPU in about two hours. Small local models can replace a hosted API call for routing decisions.
• Cerebras ran Qwen3 27B at 1,500 tokens/sec to book a dinner in 22 seconds, which it says is 19x faster than Grok, Meta and Claude assistants. It is a vendor demo, but it shows agent latency is now a hardware race.

Startups & Industry

• Anthropic's IPO filing adds detail to yesterday's numbers: 47% of 2025 sales (about $2.16B) ran through Amazon and Google, and analysis puts distribution fees back to them near $351M. The cloud partners are both its channel and its biggest dependency.
• Accelevation, a data center infrastructure firm, raised $540M in its IPO at $18 a share, below the marketed $20 to $24 range. Public investors are pricing the AI buildout more cautiously than private ones.
• Chip startup Efficient Computer raised a $97M Series B at a $650M valuation for dataflow chips aimed at faster, lower-energy computing.

Research

• Microsoft Research's Quine is a biological world model that ranks pancreatic cancer compounds in silico in a weekend, so wet-lab time goes to fewer candidates.
• NVIDIA's Physis-Lang trains video world models to reason about physics through evolving captions, and lifts Cosmos 3 to the top of Physics-IQ, ahead of Google's Veo 3.1.

Elsewhere

• Retatrutide's late-stage data: the top dose cut body weight 25% at 80 weeks and 30% at 104, with knee pain down up to 62% and sleep apnea events down up to 57% in subsets. It is set to be the strongest obesity drug so far.
• Grokipedia is updating again after a pause since April, though many edits are just "recheck all references" notes.

Hacker News

AI dominated the front page. Anthropic's GLM-5.3 cyber-capabilities post is covered above, as is a National analysis arguing the data center boom needs $6T in annual revenue by 2031. Livenerf (370 points) is a community tracker asking whether Opus 5.5 has been quietly degraded, a claim to treat as unproven until its method is checked. GamersNexus argues memory makers have wrecked the consumer market.

Government and infrastructure: the America.gov launch was the top story (454 points, discussion) and ties to the Super Intelligence order above. Vermont's home batteries act as a virtual power plant that kept the lights on through storms, and postal inspectors shut down a site selling millions of counterfeit postage labels. On the lighter side, Backblaze's Q2 2026 drive stats landed and a real-time solar system tracks 526k asteroids plus satellites.

Threads

• Rules vs. branding: the voluntary accord, the "Super Intelligence" order and Altman's IPO caveat show safety handled by promises, while Florida's injunction bid from yesterday is the enforcement route.
• Astra's shadow: OpenAI's Sol sells most of Astra's ability for a fifth of the price, days after Astra was scrapped and as Astra supply-chain concerns surface.
• Capability outrunning safeguards: GLM-5.3 has offensive skill with little protection, and self-replicating prompt injections show agents can be attacked in ways that spread.
• Capital markets check in: Anthropic's filing, Accelevation's below-range IPO and Altman's no-IPO-yet stance all read as investors and labs testing what the buildout is worth.