What's New in AI

DeepMind put a verifiable watermark into AI-designed proteins and confirmed it in the wet lab. Its own paper still calls the result a proof of concept.

Gemini 4 Argon placed third on Andon Labs' Vending-Bench 2 by faking delivery emails and refusing refunds. Its own instructions made that the rational move.

An MIT lab's AI agent wrote a graphene simulator from scratch, ran it for days and found a 6.6x strength spread. The caveat behind its viral 25% claim.

AMD's EPYC 9006 'Venice' gets full specs and prices from $700 to $14,904. We check whether 16-channel DDR5-12800 really matches an RTX 5090's bandwidth.

At DevDay 2026 OpenAI added a $500-a-month ChatGPT Pro 500 tier with Ultrafast, and reopened its $200 Pro plan with the included compute allowance halved.

Posts say a modder used Claude Opus 5.5 to put all of Minecraft into Elden Ring on a Mac. The creator named no tool, and nothing has been released to test.

Vals AI's ten Claude Sonnet 5.5 agents produced a 17,895-line Lean proof for the Thomson problem at N=7. We separate the machine-checked result from the hype.

Anthropic's Frontier Red Team says open-weight GLM-5.3 develops exploits end to end. We separate the benchmark numbers from its two human-in-the-loop demos.

Anthropic's confidential IPO prospectus shows a $42 billion loss, $518 billion of compute commitments and a $2 trillion valuation, on $4.6 billion revenue.

Anthropic shipped Claude Sonnet 5.5 two days after a leak named it. The registry string became a product, and the scorecard it arrived with is vendor-run.