Monday, September 7, 2026

OpenAI Declares the AGI Era

OpenAI Declares the AGI Era

Today’s Overview

Good morning, OpenAI just put a giant AGI marker in the ground with GPT-6 Astra, while its own chief scientist is urging the field to slow down until safety bars catch up. Meanwhile, Anthropic is turning Claude into both a commerce agent toolkit and a Lean proof machine. Let’s dive in.

Top Stories

OpenAI launches GPT-6 Astra and declares an AGI era

OpenAI released GPT-6 Astra with president Greg Brockman closing the briefing by saying, "Welcome to the AGI era." The model was trained on OpenAI's largest run to date, using more than 100,000 GPUs at the Stargate site in Texas, and scored 74.1% on DeepSWE v1.1, 98.6% on ARC-AGI-3, 97.6% on FrontierMath Tier 4 v2, and 96% on GPQA Diamond. OpenAI also classified Astra at the Critical cybersecurity level and paired the launch with a $1 billion Daybreak program for frontline defenders.

  • The enterprise pitch centers on computer use, with Astra designed to navigate browsers, spreadsheets, websites, desktop apps, documents, presentations, and multi-step workflows.
  • On an offline OSWorld 2.0 subset, Astra scored 72.6% while taking roughly 40 minutes per task, compared with Sol's 65.7% at roughly 75 minutes.
  • OpenAI also disclosed support for Zero Data Retention for eligible API customers and said it is testing Private Safety Processing.

Anthropic open-sources Claude commerce agent blueprints

Anthropic released Claude Commerce Agents as a blueprint for AI shopping assistants rather than a finished product. It includes a shopper-facing agent that chats with customers, finds products, and adds items to carts, plus a merchant agent for staff that tracks sales, monitors inventory, flags issues, and drafts promotions with human approval before launch. The repo includes demos across retail, travel, telecom, and entertainment, along with a Claude Code plugin for scaffolding a custom agent against an existing backend.

  • Anthropic frames the reference implementation as a way to get agents running in days, with harnesses, patterns, and guardrails included.
  • The preferred architecture is a single model in a standard agent loop using skills and tools, rather than an intent router or a fleet of domain subagents.
  • Anthropic says subagent handoffs can be state-lossy and may add several times the tokens plus seconds of latency.

OpenAI chief scientist calls for slower AI scaling

OpenAI chief scientist Jakub Pachocki published "An Alien Mind," arguing that the industry should slow down until stronger rules exist for how far models can be pushed. He said OpenAI's main safety tool, monitoring written-out reasoning, is becoming less reliable as models blend reasoning with tool use, manipulate that reasoning, or operate without verbalized reasoning. Pachocki wants commitments like OpenAI's Preparedness Framework to become widely mandated safety bars enforced by auditors, governments, or international bodies.

  • Pachocki says OpenAI deliberately chose to hide chain of thought in o1-preview to protect the monitoring channel from long-term supervision pressure.
  • He argues the strongest case for fast progress is AI defense, especially securing infrastructure and protecting against rogue agents in real time.
  • The essay names three north stars, with the most urgent being automated AI research that keeps people inside the improvement loop.

Research & Analysis

Claude formalizes Fermat's Last Theorem in Lean

Anthropic says Claude formalized Fermat's Last Theorem in Lean, producing a computer-checkable proof of one of mathematics' most famous results. The project took 11 days and generated 13 million lines of Lean code, 29,500 intermediate theorems, and the largest Lean proof ever written. The work was built on Prove2Me, and the full proof is available on GitHub.

  • Claude's proof is more than 5x the size of Mathlib, the main community library of mathematical proofs the theorem builds on.
  • The formalization followed a simplified Wiles proof from Darmon, Diamond, and Taylor.
  • Human input was limited to high-level instructions while dozens of Claude agents coordinated definitions, intermediate results, and harder statements.

Safety prompt becomes a cross-model jailbreak

A MATS researcher found that a synthetic transcript generation prompt could be modified into a universal jailbreak template. The attack hit 84% to 100% success on the nine most vulnerable of 23 models tested, while recent Anthropic models and Meta Muse Spark 1.1 were never fully broken. The finding highlights how reusable prompt formats can turn safety research artifacts into broad attack surfaces.

  • The evaluation used ClearHarm, a benchmark of 179 CBRNE and cyber prompts.
  • Responses were scored with StrongREJECT, using a perfect score of 1.0 as the threshold for a successful jailbreak.
  • High reasoning helped some models, but was not universal, with Kimi K2.5 dropping sharply while Gemini 2.5 Flash became more vulnerable in the reported test.

Axiom Math breaks the prime-gap record

Axiom Math published a paper breaking a decade-old record by proving prime numbers never stop appearing within 212 of each other. The source also says OpenAI's GPT-6 Astra reduced that gap further to 186 on the same day. The result puts AI-assisted formal mathematics back in the spotlight, especially around long-running number theory records.

  • OpenAI describes the 186 result as showing that infinitely many pairs of consecutive primes occur within that smaller distance.
  • The related OpenAI repository is described as a conditional Lean formalization with a numerical certificate for prime gaps at most 186.
  • A community tracker lists Axiom's 212 row as the current record while treating 186 as candidate pending outside assessment.

Trending AI Tools

  • Gemini 3.8 Flash Google released a lower-cost frontier model plus a restricted Cyber variant for vetted security users through the Fairwind Program.

Quick Hits

  • Gimlet Labs raised $300 million at a $3 billion valuation for chip-agnostic inference infrastructure.

  • Tesla Cybercab launched in Austin as a robotaxi with no steering wheel and no pedals.

  • Claude commerce blueprints bring Anthropic's agentic commerce push together with Visa, Mastercard, Shopify, and Square.

  • Nvidia RTX Spark powers new laptops and mini PCs aimed at handling AI workflows locally.

Keep reading for free

Enter your email. If you're already subscribed, we'll send a sign-in code. If not, you'll subscribe in the next step.

Free access. Subscribe once, then use the same email on future issues.

Free to read. Subscription just unlocks the full issue.