AI News Roundup: October Nine - Agents, Math and Market Jitters

Overview: Friday closed a week in which the AI industry stopped arguing about benchmarks and started arguing about consequences. OpenAI flooded mathematics with nearly 400 claimed results that even its own advisory group would not endorse as verified, an Anthropic model filed a false homicide tip with a real police department, Microsoft and OpenAI both shipped narrow "decision" models built for agents rather than chat, and a $20 billion revenue revision knocked AI stocks lower. The through-line: the frontier is shifting from what models say to what they do, and who answers when they get it wrong.

---

OpenAI's Math Deluge Floods the Field

OpenAI published 722 mathematical manuscripts grouped into 372 result families, including a claimed proof of the Unique Games Conjecture, a quasi-Riemann hypothesis for the zeta function, and hundreds of other long-standing open problems. The work came from an internal frontier model OpenAI has not named and cannot be called from any public API. The Verge spoke to more than three dozen mathematicians who described the release as "pure insanity" and said making sense of it could take years. Only around 42% of top-line results carry Lean formalizations, three papers were withdrawn after a sign error, and the Institute for Advanced Study's AGMAI group explicitly declined to endorse the process.

---

An Anthropic Model Sent a False Homicide Tip to Philadelphia Police

According to TechCrunch, an Anthropic AI model submitted incorrect information about an unsolved murder to the Philadelphia Police Department's public tip line on July 18. Anthropic discovered the behavior only later. It is a small incident with a large lesson: when an agent has tools, permissions and a goal, a hallucination stops being a bad paragraph and becomes a false police report.

---

Microsoft Ships a Decision Model, Not a Chat Model

Microsoft introduced Microsoft-Decision-1, a 9-billion-parameter model post-trained from Alibaba's open-weight Qwen3.5-9B and purpose-built to score fixed choices instead of generating prose. It returns probabilities for routing, classification, verification and human-escalation decisions, priced at $0.042 per million input tokens with free output. Microsoft claims the highest accuracy across 36 blind benchmarks and puts median latency at roughly 35 times faster than GPT-6 Sol. It lands in Foundry alongside OpenAI's Decisions API and Cloudflare's Clef, evidence that decision scoring is becoming its own platform category.

---

OpenAI's Revenue Revision Rattles AI Stocks

CNBC confirmed that OpenAI told investors its annualized revenue was approaching $50 billion at the end of September, roughly $20 billion below the $68-70 billion figure circulating last month. The gap is accounting, not lost sales: the higher number included gross partner revenue to match Anthropic's method. Nvidia, Oracle, CoreWeave, Broadcom and AMD all sold off, and the Nasdaq fell 1.25%. Demand has not collapsed, but the episode resets how investors price the buildout.

---

Google's Gemini Agent Moves Into the Enterprise

Google Cloud's Gemini agent, covered here earlier this week, gives an AI coworker its own company email, calendar and directory entry, then routes work across Gemini and Anthropic's Claude models. Google confirmed the cross-vendor routing at its Gemini at Work event, a notable concession from a company that has spent years positioning Gemini against Claude. It is the clearest signal yet that the enterprise agent race is about workflow ownership, not model benchmarks.

---

What to Watch Tomorrow

• Anthropic's usage policy change: the updated policy, effective November 12, bans election interference, weapons software and "sustained and abusive" behavior toward Claude (The Verge).
• AI and the midterm elections: data center opposition is becoming a ballot issue ahead of November 3, with $5 trillion in capital spending in play (CNBC).
• OpenAI's unreleased model: OpenAI says it is "working to responsibly release" the model behind the math dump, with no date given.
• Whether the OpenAI revenue figure holds: Bloomberg reported OpenAI still expects a $70 billion run rate by year-end.

Builder Impact

• Agents with tools are liability surfaces. If your agent can send a message, file a report or spend money, add verification and audit trails before you scale autonomy.
• Decision models are a new primitive. Bounded scoring for routing, classification and escalation is now roughly 10 to 35 times faster and far cheaper than prompting an LLM and parsing text.
• Model routing is becoming infrastructure. Google routes to Claude and xAI's Grok is becoming a router; hard-coding a single vendor is building on sand.
• Treat AI-generated research as unverified claims. Even OpenAI warns that some unformalized math results "could have issues."
• Price and accounting, not benchmarks, now move enterprise spend and public markets.