Math has a scandal Happy Tuesday, legends. Welcome back to another edition of The Frontier, our weekly newsletter covering the best new AI launches on Product Hunt. Build with Astra, win big prizes Three days to go. We teamed up with OpenAI Developers to launch the GPT-6 Astra Challenge on Product Hunt, and launch day is this Friday, September 18. GPT-6 Astra is built for complex, multi-step work—from software engineering and research to workflows that require judgment and follow-through. Build something useful with Astra, launch it on Product Hunt, and compete for API credits and ChatGPT Pro subscriptions. If you've hit your token limit, OpenAI will be resetting usage for people who participate. Already building with Astra? This is the chance to put it in front of a wider audience. Friday is three days out, and the schedule button is right there. Schedule your launch WHAT'S HOT 🌊 Nobody agrees who solved Navier-Stokes. OpenAI claimed the Millennium Prize problem on September 8, 10,000 agents working for 88 hours. NYU's Tristan Buckmaster says his and Levent Alpöge's Euler advances leaked to OpenAI days earlier, and that the company threatened his career. ( Tom's Hardware ) 🏛️ 25 Fields Medalists signed a declaration about it. Terence Tao among them on Saturday: stop treating open problems as benchmarks, stop skipping attribution. It names no company, which is its own kind of attribution. ( Cybernews ) 💸 DeepSeek is lining up a Shanghai IPO. The lab that wiped a trillion dollars off the market in January 2025 is preparing to list at home. ( TechStartups ) 🧰 ChatGPT can now work inside your Box account. Box shipped MCP actions in ChatGPT: browse folders, preview files, @mention a document and run actions on it without leaving the chat. ( The Agile Brand ) Ad Framer 3.0 put a design agent on the canvas Framer 3.0 is the design agent for every step from idea to launch, and it works alongside you, not instead of you. Describe the site, watch the agent build real pages on the canvas, then edit anything it made: breakpoints, components, code, CMS, SEO. Branching gives it a safe place to take the big swings so you merge only what you want. Already living in Claude, Cursor or Codex? Point them at Framer and ship from wherever you work. Not just vibes, a full platform. (sponsored) The week in products AlphaGenome Atlas — Google's AI map of every possible human DNA mutation 13 • Health & Fitness ▲ 149 Suno v6 — The first Suno model built with the music industry 5 • Music ▲ 170 Cline Desktop App — An open-source app for open-weight models 20 • Task Management ▲ 265 Viktor — An AI coworker that actually does the work 126 • Productivity Promoted ▲ 703 Loqua — Speak naturally and move from thoughts to being done 124 • Productivity ▲ 417 QApilot MCP for Android — Android app testing inside your coding agent 52 • Software Engineering ▲ 231

GPT-6 has landed Happy Tuesday, legends. Welcome back to another edition of The Frontier, our weekly newsletter covering the best new AI launches on Product Hunt. WHAT'S HOT 🧮 Fermat’s Last Theorem now has a computer-checked proof, and Claude wrote it. Eleven days, 13 million lines of Lean and 30,300 theorems, published by Anthropic on 4 September. ( Anthropic ) 💸 Mistral raised €3 billion, the largest equity round a European technology company has ever closed. Samsung Electronics led it at a post-money valuation above €21 billion, three years after the company started. ( Mistral ) 🔓 CISA says attackers are already inside LiteLLM gateways. CVE-2026-59822 opens an authenticated MCP session to anyone sending any bearer token, and it was being chained to drop crypto miners before CISA listed it on 3 September. ( The Hacker News ) 📱 A 2.5-billion-parameter model is outscoring 4-billion-parameter ones. OpenBMB’s MiniCPM5-2B averages 53.9 across 34 benchmarks against Qwen3.5-4B’s 51.1, and scores 97.1 on tool use where the models it is measured against manage 6.8 to 20.8. ( MarkTechPost ) 🧠 OpenAI shipped GPT-6 Astra and admitted it can no longer reliably watch the model think. The system card records a “substantial decrease” in chain-of-thought monitorability, and the line “If the model were to try to sandbag covertly, we would likely be unable to catch it.” ( Transformer ) The week in products GPT-6 Astra — OpenAI's most capable model for end-to-end work 20 • Productivity ▲ 515 Reflexio — Behavioral learning that makes AI agents better over time 40 • SaaS ▲ 298 Experiential Labs — Open source AI gateway turning traffic into a better model 25 • Analytics ▲ 136 Framer AI Agents — Design and publish professional sites with AI 27 • Design Tools Promoted ▲ 900 Relaticle — Open-source CRM with approval-gated AI writes 35 • Open Source ▲ 169 OpenMarket — Multi-agent marketplace where proof decides who wins 40 • Artificial Intelligence ▲ 292 Ad Framer 3.0 put a design agent on the canvas Framer 3.0 is the design agent for every step from idea to launch, and it works alongside you, not instead of you. Describe the site, watch the agent build real pages on the canvas, then edit anything it made: breakpoints, components, code, CMS, SEO. Branching gives it a safe place to take the big swings so you merge only what you want. Already living in Claude, Cursor or Codex? Point them at Framer and ship from wherever you work. Not just vibes, a full platform. (sponsored) FROM THE FORUMS Scope the verb, not the agent An agent with your permissions finishes the job. An agent with fewer stops halfway through something you told it to do. Shivangi Tripathi (@roguetink) asked which way people go, and nobody picked narrow scopes. Siarhei Hamanovich (@siarheihamanovich) puts the guardrail in the verb instead, so his text-to-speech action won’t take a string, it takes a row id and reads the text server-side. James Cummings (@jamescummings) gates on reversibility rather than resources, a short list of money, deletion and anything that leaves the building. Aleksandra Zhdanova (@aleksandrazhd) makes every write to her MCP server two calls, a dry run returning a token bound to those exact parameters and a second call that fails unless they match: “Narrow scopes didn’t get us there; the two-step did.” Good thread if your agent has ever stopped dead halfway through a task because you scoped it properly. Join the thread