The Polylog AI Intelligence Brief

Morning Edition · Friday, November 20, 2026

Markets

Anthropic raises $81.3 billion at $83 a share in the largest IPO on record and closes its first day 17% higher

The listing puts a public price on a company whose amended prospectus shows about $900.9 billion in compute commitments, against a revenue run-rate of about $81 billion.

2 sources
Corroborated

Tech

OpenAI ships GPT-6.2 Astra and Sol with a 'Critical' cybersecurity rating and a doubled failure rate on a deception test

According to its system card, Astra is also the first public OpenAI release rated High for AI self-improvement under the company's Preparedness Framework.

1 source
Plausible

Tech

Google prices Gemini 4 Neon at $3 and $15 per million tokens, 25% below what Gemini 4 Carbon will cost from December 1

Early third-party tests put Neon one point above Carbon on the Artificial Analysis Intelligence Index, at about half of Carbon's cost per task.

1 source

Tech

AMD closes its purchase of Taalas, a Toronto startup that builds model weights directly into silicon

AMD has not disclosed the price, which one report puts at up to $2.9 billion, including $1.4 billion in contingent payments.

1 source
Corroborated

Tech

OpenAI publishes a supervised-deployment blueprint for Daybreak Red while Anthropic keeps Mythos 5.2 behind vetted access

Both labs now ship their most permissive cyber models to approved defenders through account controls, while their general releases carry the same base capability with more safeguards.

5 sources
Plausible

Tech

Microsoft opens Nvidia Vera Rubin NVL72 systems on Azure in preview and adds serverless post-training to Foundry

At Ignite, the companies also extended Azure Local to RTX PRO 6000 Blackwell and Jetson Thor hardware as Windows desktops built on Nvidia's RTX Spark reached the market.

2 sources

Tech

Two new papers show how authority leaks when AI agents sign payments and hand tasks to other agents

The research arrives as OpenAI chief executive Sam Altman reportedly told a New York Times conference that handing control to AI is "a terrible idea."

3 sources
Corroborated

Tech

OpenSSL fixes an encoder overflow rated 9.8 for severity, while its exploit-probability score stands at 0.14%

A new arXiv paper compares the two main scoring systems at a fixed analyst budget, which is the decision the gap between them forces on security teams.

3 sources

Tech

Google adds Tool Retrieval to the Gemini Interactions API so agents load tool schemas only when they need them

The preview works with Gemini 3.6 and later models, carries no separate charge, and is not available through the older generateContent method.

1 source

Tech

Cognition launches Devin Bots, persistent agents that own a goal and direct teams of Devin instances

OpenAI separately described a similar pattern of standing responsibility internally, with Codex-powered workflows that find and fix ChatGPT performance problems.

2 sources

Tech

New papers put AI agents into self-driving labs and algorithm research, with human authorization built into the design

The work extends a pattern already running at Lawrence Berkeley National Laboratory, where Meta's open vision models segment X-ray scans while experiments are still in progress.

3 sources

Tech

Google opens its Genie 4 world model to AI Ultra subscribers through Project Genie

The release adds real-time audio and prompt editing during a session, but DeepMind's own model page limits continuous interaction to a few minutes.

1 source