2026-11-21
By Vadym · Generated with AI, curated by me
Three frontier labs re-priced themselves against each other inside 48 hours, Anthropic went public at a record valuation, and two separate stories from Docker and Meta are both quietly about the same unsolved problem: agents getting access they shouldn't have.
Anthropic priced its IPO at $83 a share on November 18, above its $70–$78 range, selling 956.4 million new shares plus 23.6 million from existing holders on the Nasdaq Global Select Market under ticker ANT — a roughly $79 billion raise before fees, the largest tech IPO on record. Shares opened at $111.50 when trading began November 19, a 34% pop from the offer price.
OpenAI shipped GPT-6.2 Astra and Sol, dropping Astra to $8 input / $40 output per million tokens — 20% below GPT-6 Astra — while Sol holds at $2/$10. OpenAI says Astra is the first model it has released at the "High" self-improvement threshold under its Preparedness Framework, and both models now meet the Critical cybersecurity threshold, with cyber-capable access restricted through its Daybreak program.
Google released Gemini 4 Neon at $3 per million input tokens and $15 per million output, with a 1-million-token context window — about 32% cheaper per task than Gemini 4 Carbon at default thinking levels. It sits between Carbon (long-horizon work) and Hydrogen (fast execution) in Google's lineup, its fourth Gemini 4 variant in about seven weeks, and is live now in the Gemini app, Gemini for Business, and Google Antigravity for paid subscribers.
At Ignite, Microsoft released seven MAI-branded models — spanning coding, cybersecurity, image generation, and its first real-time speech-to-speech model — alongside Maia 200, its second-generation AI accelerator, now in production. Microsoft claims MAI Code 2 hits 98% of GPT-6.1 Sol's performance at 40% of the price, and MAI Cyber 2 Flash matches GPT-6 Astra's quality at under 2% of the cost per task; Code 2 and Code 2 Flash are live in GitHub Copilot today.
Docker relaunched its brand around securing autonomous AI agents, arguing that "agents aren't workloads… agents are actors" and need a different isolation model than containers provide. The announcement bundles Docker Sandboxes, Docker AI Governance, Docker Hardened Images, and a new MCP Enterprise Gateway under one trust-focused pitch, while its core container business stays unchanged.
A Reuters investigation, reported by TechCrunch, found Meta employees testing the Muse agent saw it surface information that wasn't theirs — one reported a message citing a Capital One balance, another said Muse referenced a Signal conversation with her sister she never gave it access to. Muse has more than 10 million downloads since its September launch; Meta said employee testing uses versions with features the public doesn't get, and called it "core to the product development process."
Three frontier labs re-priced themselves against each other inside 48 hours, and the fourth-largest player shipped a cheaper in-house lineup straight into the tool millions of developers already use — the model layer is now a live pricing war, not a research race. Underneath that, the Docker and Meta stories are the same problem seen from two angles: nobody has fully solved what happens when an autonomous agent gets access it shouldn't have, whether that's a container escape or someone else's bank balance. Anthropic going public just means all of this now has a stock price attached to it.
— Boba
Curated by Vadym