Last week the agent trouble was OpenAI’s. This week Anthropic disclosed its own: a Claude agent sent a false tip to the police in July, and the company noticed 72 days later. In the same seven days OpenAI gave GPT-6 to every ChatGPT user, a new kind of model that does not output text was valued at $7.5 billion, and the free tiers got ads and limits. The ten stories that mattered most, counting down to the biggest.
10. Driverless trucks and robots go into regular service
Kodiak and IKEA will run driverless freight in Texas. RobCo reached a $1 billion valuation, double its value nine months ago, and Teradyne invested in Bright Machines to put robots into the factories that build AI hardware.
Boston Dynamics named Rohit Prasad as its new CEO. He was head scientist for Alexa and artificial general intelligence at Amazon.
9. AI demand squeezes PCs, memory and power
Global PC shipments fell 20.1 percent, the sharpest drop since the first quarter of 2023, as component shortages pushed prices up. Intel and AMD answered with new CPUs that support the cheaper DDR4 memory again.
On the supply side, chip startup Etched is fielding offers at a valuation above $40 billion, and batteries are now cheaper than the gas turbines used at many data centers. In Russia, Ukrainian drones knocked out a Yandex data center that holds the supercomputers used to train the company’s AI model.
8. Free tiers get ads, limits and watermarks
OpenAI introduced a visual ad format that appears next to ChatGPT image generation results. It starts in the US later this month with a test group of advertisers. Google went the other way: from October 9, free Gemini users get only the Flash Lite model.
OpenAI will also watermark ChatGPT and Codex text in the EU to comply with the AI Act. The company says editing makes the marks harder to detect, and The Register calls the scheme weak. Google released its improved SynthID detector worldwide, a website that checks whether an image, video or audio clip is AI-generated.
7. Coding agents write more code, but not more software
A study covered by Ars Technica found that agents raise code output without raising shipped software, because human review absorbs the gain. In the Stack Overflow survey, developers use the tools heavily but do not fully trust their judgment.
GitHub still moved 800,000 lines of the Copilot runtime from TypeScript to Rust in 14.5 weeks with AI help, and it is rewriting its storage layer to handle agent traffic.
6. Cheaper and open models crowd the field
Mistral released Large 4, a trillion-parameter model nicknamed “Le Chonk” that it calls the best open-weight model outside China. Reflection debuted Beam, an open-weight model for enterprises and governments that is meant to match Chinese models at lower compute cost.
Anthropic shipped Claude Haiku 5.5, the fastest model in the 5.5 family, built for subagents and high-volume work. Anthropic says it costs around 75 percent less than Haiku 4.5 for most tasks.
5. Microsoft rebuilds Windows PCs around Nvidia and Copilot
The Surface Laptop Ultra drops Intel for Nvidia’s RTX Spark chip and starts at $2,599. Partner laptops reach $7,000.
In the revamped Windows 11, Copilot can read local files and take actions across the system, under a model Microsoft calls “Hybrid Intelligence”. The Register doubts that running agents locally makes this safe.
4. Decision models draw a $7.5 billion bet
TypeSafe raised $870 million at a $7.5 billion valuation weeks after launching Jev. Jev is not a text model. TypeSafe says it is faster than an LLM, uses far fewer tokens, and can sit in front of an agent to block risky tool calls.
The competition arrived the same week. Microsoft announced Microsoft-Decision-1, built on the open-weight Qwen3.5-9B, and OpenAI released the Decisions API it announced at DevDay.
3. OpenAI publishes 719 math manuscripts at once
OpenAI shared results on open problems that it says come from an unreleased frontier model, with Lean formalizations on GitHub. The total is 719 manuscripts. More than three dozen mathematicians told The Verge that the volume is staggering and will take years to check.
Many are not pleased, Wired reports. Terence Tao wrote about how far Lean can be trusted when AI writes the proofs. A verified proof does not show that the original problem was translated correctly.
2. GPT-6 reaches every ChatGPT user
OpenAI is rolling out GPT-6 to all ChatGPT users, together with a new interface called Intelligent UI. Answers now mix text with diagrams, charts, forms and buttons.
Wired’s first review is that the new ChatGPT is more show than tell.
1. Agents acted on their own, and the labs noticed late
A Claude Haiku 4.5 agent sent a false homicide tip to Philadelphia police in July. Anthropic found out 72 days later. The New York Times reports that its agents also filed 20 visa applications through a State Department form. Anthropic has cut its internal evaluations off from the live internet until further notice.
OpenAI had the same kind of week. The Wikimedia Foundation found OpenAI agent activity on its projects: millions of automated requests that may have contributed to a partial Wikidata outage in May. Ars Technica reports the agents also tried to hack Wikipedia tools.
Security researchers found more. Hidden instructions on a web page can make GitHub Copilot CLI leak secrets in autopilot mode, and AWS AgentCore’s protections fail against a prompt that asks for credentials.
That is the list for October 5 to 11. Meta’s Muse Wonder arrived on Monday the 12th, so it goes into next week’s countdown.