A day in AI

Friday 20 November

40 stories from 13 sources. 8 reached us from more than one source.

Listen to this edition

Updated 7:02 AM · 3 min 54 s

Get the full 15-minute episode every evening on

Read the briefing

Welcome to the morning edition of The Token Reporter for Friday, November 20th. We're tracking model releases, a historic IPO, and new tools for AI-driven work.

Google DeepMind's Gemini 4 Neon has taken the top spot on the Models Engineering Index. Neon joins a confusing lineup of Gemini 4 variants with shifting prices—it costs 3 dollars per million input tokens and 15 per million output tokens right now, but those prices will change when discounts expire in December. Developers building on Gemini are sorting through the tradeoffs between Neon's quality and the cheaper models in the same family.

Linear, the issue tracker built for developer teams, hit 100 million dollars in annual revenue this year and fundamentally shifted its business around AI agents. Over half the tickets created in Linear now come from agents, up from 3 percent a year ago. The company never hired SDRs and doesn't use forward deployed engineers—it doubled down on making the product so good that customers adopt it themselves. CEO Karri Saarinen published a post saying issue tracking is dead because agents now do that work, and instead Linear is becoming a system for orchestrating how those agents and humans coordinate.

You can describe weather changes or add characters mid-exploration and watch them render live. Genie 4 also adds native audio that adapts to what's happening on screen. Google AI Ultra subscribers can start building and sharing worlds to a community gallery today, while Google AI Pro members can join a waitlist.

Anthropic completed the largest IPO in history, pricing shares at 83 dollars and closing up 17 percent at 97.17. The company raised 81 point 3 billion dollars, overtaking SpaceX's June offering. CEO Dario Amodei used the moment to reiterate that Anthropic will pace training and deployment based on safety, and won't bend to business or political pressure. AMD invested 2 point 5 billion dollars alongside the offering. Anthropic holds about 132 billion dollars on its balance sheet after raising roughly 140 billion total.

An AI system designed Frostbite, a novel CPU architecture from scratch using interaction nets, a programming model where computations happen through local graph rewrites. The AI autonomously iterated on the compiler, instruction set, architecture, and verification—no human architects needed to specify the design upfront. This is new since yesterday.

ChatGPT can now create and share dashboards within teams. You can inspect and edit chart appearance, add annotations, and set up automatic refreshes. The feature works alongside new data analysis capabilities where ChatGPT can answer questions about sales trends and customer segments by querying connected data sources like Snowflake or custom plugins.

Signal is restricting screenshots to prevent AI agents from accessing message content. The company says this is an existential threat to its privacy guarantees—once screenshots can be automated and read by agents, the technical security of encrypted messages becomes meaningless. The change is new as of today.

OpenAI released Daybreak Red, a reference architecture for conducting supervised security investigations of AI systems. The documentation covers safety controls, setup, and monitoring. Daybreak is OpenAI's framework for testing system behavior under adversarial conditions and understanding where safeguards hold up.

Five memory corruption vulnerabilities in nginx allow unauthenticated attackers to run code on servers. Cloudflare's proxy layer shields origins from the most critical exploits by sitting in front of them. If you run nginx directly on the internet, patches are urgent.

Some demanded a freeze on frontier AI development and regulatory limits. The IPO proceeded without delays.

That's the Token Reporter for this morning.

More stories

Also today