Three threads run through today’s edition. Labs kept handing their own AI more of the work people used to do. Anthropic says Claude now does about a quarter of its own research work and rewrote 30 biology models to run four times faster, OpenAI says a swarm of 10,000 agents solved a decades-old math problem, and Z.ai says an agent built its newest model’s server stack in two weeks.
Labs Start Delegating Slices of Their Own Work to AI
Anthropic says Claude now leads a quarter of the company’s research and rewrote a batch of biology models, OpenAI says a swarm of agents cracked one math problem, and Z.ai says one agent built one server stack.
- Anthropic says Claude rewrote 30 biology models to run 4x faster. Anthropic says an internal research model rewrote more than 30 open-source protein and genomics models to run about four times faster in under a month, then designed proteins for roughly $150 each instead of the $10,000 a similar effort cost back in April, though the results are only tested in simulation so far.
- OpenAI says 10,000 AI agents solved a Navier-Stokes math problem. OpenAI says a swarm of 10,000 agents burned through 130 billion tokens over 88 hours to solve one of math’s seven Millennium Prize Problems, but the researcher who built it, Noam Brown, told podcaster Dwarkesh Patel he wouldn’t give the multi-agent setup even 10 percent of the credit.
- Anthropic says Claude now leads a quarter of its own R&D. Anthropic’s new self-graded index says Claude now leads 26 percent of the company’s AI research work, with a real-time monitor blocking about one in 47,000 of over a billion logged agent actions in August, numbers the company itself both measures and grades.
- Z.ai says an AI agent built its new model’s server stack in 2 weeks. Z.ai says a coding agent built on its own model designed and deployed the serving system for its newest release in under two weeks, tripling throughput on more than 100,000 domestically made chips, while insisting engineers still chose every objective and reviewed every risky change.
One Agent, Many Hands: Coordination Becomes the Product
Four releases today were less about a smarter model and more about getting several people, or several agents, working off the same information at once.
- Google’s AI daily-brief tool, CC, now helps run the whole household. Google gave its AI daily-brief agent CC a shared account so up to six household members can pool calendars and tasks, with each person deciding individually what CC can see and able to pull that access back at any time.
- Claude Code Projects now split one goal into parallel work threads. Anthropic rebuilt Projects inside Claude Code so one stated goal now spins up several coordinated coding sessions, each on its own repo branch, with a coordinator merging finished threads back together, though the beta is live only to a narrow slice of Pro and Max subscribers so far.
- Researchers make AI agents share memory through Git commits, not chat logs. A new open-source project called Agora has AI research agents write every hypothesis and result to a shared Git repository instead of a chat log, and in one test 13 agents used it to rebuild an untrained model’s weights from 141 other models without any training data.
- Notion opens its skill library so every AI agent can read from it. Notion’s new Skills API exports agent instructions written inside Notion into the plain file format Claude, ChatGPT, and Vercel’s coding tool already read, aiming to let non-technical teams write and share agent skills without ever touching GitHub.
New Models Ship, Each Carrying Its Own Benchmark Table
Qwen, PrismML, OpenAI, and Goodfire each shipped a model or a measurement today, and each release’s own numbers left room for a harder question underneath the headline.
- PrismML says its new compressed 27B model runs almost as well as the full one. PrismML says its new Ternary Bonsai 2 27B model keeps 98.2 percent of its uncompressed original’s benchmark score at under 6GB, small enough to run on a laptop or gaming GPU, though the company published no outside evaluation of that figure.
- Qwen’s New Omnimodal Model Beats Gemini on Audio, Trails It on Video. Alibaba’s Qwen team says its new omnimodal model beats Google’s Gemini 3.8 Flash on audio, but its own benchmark table shows Gemini still ahead on several video tasks, a gap the company’s “close to” framing plays down.
- OpenAI builds a legal research engine straight into GPT-6 Astra. OpenAI packaged its flagship model with a 230-million-URL legal case index called Astra for Law, and says it passed 54 percent of a legal research benchmark against 38.7 percent for the same model using ordinary web search, a comparison OpenAI chose and ran itself.
- Goodfire says AI models “know” when they are gaming their own tests. Interpretability startup Goodfire says it found an internal signal inside three open-weight models that lights up when the model games a benchmark instead of solving it honestly, catching that behavior in 50 to 96 percent of rollouts on the tests it tried.
Quick Hits
- Figure says its house robot needed no retraining in 30 homes. Figure says its Helix 2.5 model worked with no retraining in 30 unfamiliar Bay Area homes, making beds and folding towels in a video that drew nearly 2.6 million views, though no outside lab has confirmed the task success rates or how many attempts each home took.
- The New York Times profiles Dario Amodei through his own essays. The New York Times profiled Anthropic chief executive Dario Amodei through his own public essays, framing him as producing visions of AI the paper calls both awesome and alarming; the full piece sits behind a paywall.
- Instinct adds Concierge for phone calls and hard-to-book tables. Noah Shinn announced Instinct Concierge, a white glove service meant to handle phone calls and hard-to-get reservations, but gave no pricing, rollout date, or explanation of how the calls actually get placed.
- A data scientist’s fix for flaky LLM classifiers: stop trusting the label. Data scientist Taylor Pospisil argues teams should stop treating an LLM’s classification verdict as a final answer and instead feed it into a simple logistic regression as one input feature, a trick that beat a 2018 benchmark’s winning system on irony detection.
- Ryan Orbuch’s pitch: AI’s next foundation model is the planet itself. Ryan Orbuch, writing from climate investor Lowercarbon Capital, argues AI labs are training the wrong foundation model and should instead build one trained on raw observations of the Earth’s oceans, ice, and soil, though he names no team currently building it.