Anthropic’s Biolab Claude Discovers CRISPR-Like Enzyme System The Verge — Anthropic says Claude “autonomously discovered” a new enzyme system analogous to the machinery behind the powerful gene-editing tool CRISPR. It’s the first result from the company’s newly-launched wet lab and an early test of Claude’s usefulness for science as Anthropic prepares to go public. The company describes the finding as among the most significant AI-driven biological discoveries to date.

🚀 Frontier Models

Claude Opus 5.5 — Anthropic

Anthropic launched Opus 5.5, matching Claude Fable 5.1 on most benchmarks while running at 40% lower cost than Opus 5. The model passed Anthropic’s strongest automated behavioral audit to date and was also evaluated by external auditors. A companion post breaks down what a task actually costs, showing that extensive cache reads and flexible effort levels make it highly token-efficient across session lengths.

GPT-6 Sol and Luna — OpenAI

OpenAI launched Sol and Luna as faster, more affordable siblings to GPT-6 Astra, priced at roughly half the equivalent GPT-5.6 tier. Both models advance coding, factuality, computer use, and professional tasks at lower cost. Simon Willison’s analysis of the dual launches notes OpenAI and Anthropic published within an hour of each other, making the pricing reset feel coordinated.

SWE-Bench Pro V2 — Scale AI — The harder successor to SWE-Bench drops top models (including GPT-5 and Claude Opus 4.1) to roughly 23% on 642 real-world tasks, correcting prior evaluation errors and adding multi-file complexity.

Better GPT-6 Prompt Caching — OpenAI — OpenAI raised default cache hit rates and added 30-minute shared-prefix discounts alongside new monitoring tools for diagnosing cache performance.

⚖️ Policy & Safety

Altman and Amodei Expected to Join UN Security Council Meeting on AI — CNBC

OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei will address the UN Security Council on AI safety amid mounting international concern about AI risks. France, which holds the September presidency, called the session — the first time the Council has directly convened frontier AI developers from both the US and China together on shared safety issues.

Bernie Sanders Proposes Banning Superintelligence — The Verge

Sen. Bernie Sanders and Rep. Greg Casar introduced the Ban Artificial Superintelligence Act, which would outlaw developing AI capable of “destruction or disempowerment of humanity” — including overthrowing the government. Violators could face up to 20 years in prison. The bill defines superintelligence as a threshold, not a current technology.

Data Centers Are Black Boxes, but California Wants to Change That — The Verge — Gov. Gavin Newsom signed a package of bills requiring data centers to disclose their electricity and water consumption to local communities, responding to protests over AI infrastructure costs.

Even Americans Who Use AI Every Day Are Worried About It — TechCrunch — A new report finds heavy AI users remain anxious about the technology and broadly support regulation, suggesting familiarity alone won’t reduce public unease.

🛠️ Developer Tools & Infrastructure

Claude Marketplace: Plugins, Agents, and Services from Partners — Anthropic

Anthropic launched Claude Marketplace as a single destination to discover and connect plugins, partner agents, and third-party services. Early partners include Snowflake, Vercel, CodeRabbit, Power Digital, and ThoughtSpot. A companion post covers how to prepare codebases for AI-driven modernization.

Python Workers Are Now Generally Available — Cloudflare — Python Workers graduate from beta with native support for Cloudflare bindings (R2, D1), popular frameworks (FastAPI, Django, Flask), and AI libraries like LangChain and the OpenAI SDK.

Microsoft Open-Sources TauGrid for AI Workloads on Kubernetes — InfoQ — TauGrid unifies scheduling, queuing, and monitoring for GPU-based AI training and inference through a single Helm install, built on Kueue and KubeRay.

Google Publishes RRSI for Self-Improving AI Agents — Google — RRSI regularizes how agent harnesses recursively self-improve, reducing benchmark overfitting while improving out-of-distribution performance across eight benchmarks.

Hardware-Agnostic Models in vLLM — PyTorch Blog — vLLM’s new hardware-agnostic layers reach 96.6% of native H100 efficiency while remaining torch-compilable and extensible to non-NVIDIA accelerators.

Ax — Google (GitHub) — High-throughput declarative orchestrator for running billions of autonomous agent workloads on Kubernetes, with primitives for network fencing and cost control.

LiteParse — LlamaIndex (GitHub) — Standalone, locally-runnable OSS PDF parser with high-quality spatial text extraction and bounding boxes, no cloud dependencies.

💰 Funding & Enterprise

Snorkel AI Triples Valuation to $3.5B — TechCrunch

The AI training data platform raised a $350M Series E as enterprise demand for programmatic data labeling and data-as-a-service surges. The seven-year-old startup now counts over $500M raised in total, with the tripling valuation reflecting how central training data has become to AI strategy.

Bessemer Raises $5.75B, Targeting AI-Native Companies — TechCrunch — The VC firm closed a new fund saying AI-native companies are growing faster than any technology ever, signaling continued capital concentration in the sector.

Enveda Secures $311M for Nature-Derived AI Drugs — TechCrunch — The AI biotech, now valued at $2B, is advancing drugs targeting skin conditions and weight-loss preservation after GLP-1 discontinuation.

Ema Raises $77M as AI Eats into Enterprise Software — TechCrunch — The enterprise AI agent platform has raised $140M to date, with Google and Microsoft among its 50+ enterprise customers.

🤖 AI Products & Agents

Meta Admits Muse’s Likeness to OpenClaw Isn’t a Coincidence — TechCrunch

Meta acknowledged that its Muse AI agent draws heavy inspiration from the open-source OpenClaw project, despite claiming it was built from scratch. The admission follows a hands-on from The Verge showing Muse as a shopping agent that completed purchases autonomously.

YouTube Builds AI Tools for Creators and Custom Algorithm Feeds — TechCrunch — At Made on YouTube, the platform announced AI thumbnail testing and idea generation for Studio, plus Gemini-powered custom feeds letting users describe what they want to see in plain language.

ChatGPT Mobile Gets Voice-Based Agentic Features — TechCrunch — Pro and Plus users can now use the Work tab on phones to complete multi-step agentic tasks via voice.

OpenAI Nabs Key Patreon Execs for Creator Strategy — The Verge — Patreon co-founder Sam Yam joins OpenAI to lead Creator Product, bringing the platform’s former product and engineering heads ahead of an expected creator-focused announcement.

Rabbit’s OS3 Standalone AI Agent Drops the R1 Hardware Requirement — The Verge — Rabbit’s new agentic OS runs in the cloud but operates locally across Windows, Mac, and Linux — no R1 device needed.

Generated by claude-sonnet-4-6 on 2026-09-23T10:00:00Z