Anthropic ran an end-to-end binder design campaign with Mythos Preview and Opus 4.8. Of 1,320 designs submitted, 354 confirmed as binders in wet-lab testing by Adaptyv Bio and Twist Bioscience - a 26.8% hit rate against a 10 to 15% industry average, across 14 of 15 targets. On the RBX1 target alone, Mythos Preview hit 40% versus a 3.7% baseline from a prior open competition. The full 1,320-design dataset is public on Hugging Face, so the numbers are checkable. Read the research →

Sarah Friar told staff at an August 19 all-hands that OpenAI will be a public company in 2027 or sooner. Numbers shared: revenue run rate up 35% quarter to date, enterprise revenue up 50%, annualized run rate past $40 billion, and 20 million weekly active coding users. Source →

Jane Street is running Etched inference chips in production, disclosed alongside a $700 million raise led by Jane Street on August 18, per Etched's press release. The valuation doubled in under 30 days from the $10.3 billion Series C on July 23, bringing total raised to $1.9 billion. A quant firm putting a transformer-specific ASIC into live production is the strongest signal yet that Nvidia's grip on inference has a real challenger. Read our writeup →

Apple filed its August 19 response to OpenAI's motion to dismiss, reaffirming claims of pervasive trade-secret misappropriation by two former engineers who joined OpenAI's io Products hardware division. Oral argument is set for October 1. If the judge denies the motion, the case moves to discovery - putting OpenAI's hardware roadmap and hiring practices under subpoena. Source →

Cursor's August 19 changelog adds event-driven subscriptions: an agent can subscribe to a GitHub PR thread, a Slack channel, or a timer and activate when an event fires. An agent that opens a PR watches its own failing CI and answers review bots without re-invocation. Custom Modes pin skill sets as always-on instruction layers. No separate pricing was announced. See the changelog →

Warp's Factories beta puts pipeline definitions in YAML committed to the repo, with each stage running a coding agent (Claude Code, Codex, Cursor, or open-weight) plus optional human checkpoints across triage, spec, implementation, review, and monitoring. Qualified orgs get $10,000 in starter credits. Defining an agent pipeline as reviewable code is the missing piece for teams that want agents in CI without losing an audit trail. Read Warp's post →

The newsletter platform behind this Brief. Clean API, fair pricing, no AMP nonsense. Try Buttondown →

LLM-ready web scraping that turns any URL into clean markdown for your agents. The extraction layer the agentic browser comparison below leans on. Try Firecrawl →

Kitesurf, Browserbase, and Browser-Use solve three different agent-browser problems. Pick Kitesurf for one-shot extraction on Cloudflare Workers, Browserbase for multi-step workflows that survive bot checks, and Browser-Use when content cannot leave your own infrastructure. Read the guide →