🏆 THE AGENT REPORT — WEEKLY DIGEST

Hi AI builder,

Here are the top 5 AI Agent stories from this week:

1. OpenAI's 10,000 Agents Crack Navier-Stokes, a $1M Millennium Problem, in 88 Hours OpenAI says 10,000 autonomous agents running on an unreleased internal model found a singularity in the 3D Navier-Stokes equations — resolving one of the six remaining Millennium Prize Problems — in just 88 hours. The swarm exchanged roughly 3 million messages, and a second model spent 17 more hours formalizing the proof in Lean. It's a landmark for multi-agent reasoning, though a priority dispute with NYU's Tristan Buckmaster and Anthropic's Levent Alpöge followed within 12 hours.

2. GPT-6 Astra: OpenAI's Flagship That Finishes the Work OpenAI's new flagship landed September 3 with a 1.05M-token context window, five reasoning-effort levels, and $10/$50 per-million-token pricing — a 2.5× jump over GPT-5.6 Sol. It's tuned for autonomous work: DeepSWE v1.1 at 74.1, OSWorld 2.0 at 72.6, and a perfect ExploitBench score gated behind Trusted Access. One catch: the API defaults reasoning effort to "low" despite the "Highest reasoning" marketing.

3. The Distillation Boomerang: NSA Names Six Chinese Labs, a Humanoid Startup Names OpenAI On September 8, the NSA, CISA and FBI named six Chinese labs — DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI — for industrial-scale distillation of US frontier models. Two days later the accusation boomeranged: JoyIn, an Ant Group-backed humanoid startup, accused OpenAI of distilling its robotics model and began legal proceedings. The real story is the now-symmetric public standard of proof — and a recommended defense that silently degrades outputs for suspected accounts.

4. Gemini 3.8 Flash and Flash Cyber: Google's Cost-and-Cadence Counter to GPT-6 Google answered GPT-6 week not with a bigger flagship but with cadence and cost: Gemini 3.8 Flash, its third Flash release in six weeks, posts the best Flash-class DeepSWE v1.1 score at 73.7% for $0.75 per million input tokens. A sibling, Flash Cyber, patches CWE-Bench at 47.2% pass@1 — near frontier quality, but only for vetted defenders. It's a bet that the agent economy rewards the floor, not the ceiling.

5. AI Agent Funding Q3 2026: 20 Rounds, $1.32B, and the Agents-Replace-SaaS Thesis Q3 2026 logged twenty AI-agent funding rounds worth roughly $1.32 billion, and August's seven rounds hit $681.5M as the average cheque doubled from $49M to $97M. Capital is concentrating on infrastructure, orchestration, and the security/control plane — not the agent itself. The thesis now being funded: a SaaS company without native agentic capability will struggle to raise at any stage.

📖 Read the full coverage: https://the-agent-report.com/latest/ 💌 Subscribe: https://buttondown.com/theagentreport See you next week! — The Agent Report