Zvi's AI weekly: Gemini 4 Argon lands at $2/$10, GPT-6.1 Sol steps in after Astra alignment failure

Don't Worry About the Vase (Zvi) · rss · 2026-10-01

Key points from Zvi's weekly AI roundup:

Gemini 4 Argon is rolling out with claimed frontier-level benchmarks at $2/$10, but access is unavailable pending—Google marketing failing again.
OpenAI pulled GPT-6.1 Astra due to alignment failures and shipped GPT-6.1 Sol instead, pitched as near-Astra quality at the same $2/$10 price.
Other Dev Day news: Ultrafast mode and Dot, an always-on Astra-based agent bundled with Pro subscriptions.
The author finds Claude Opus 5.5 the best model by a wide margin, considerably boosting his productivity.
The White House hosted tech leaders and signed a 'morally binding' AI safety accord.
Jensen Huang on the Ezra Klein Podcast: treats AI safety as an engineering problem; many claims disputed.
Also: Anthropic IPO prospectus leaked; Anthropic's report on GLM-5.3 cyber capabilities; Google testing TPUs in space; California gene synthesis guardrails; hospitals using AI for aggressive upcoding; Nvidia's Open Agents Safety Platform.
Utility highlights: Anthropic employees completed N=4 super Yang-Mills to nine loops with Fable 5.1 for $100 in credits; Carter Church one-shot broke an unsolved Napoleonic cipher with Astra; Patrick McKenzie: model output went from 'median low-seniority coworker' to 'frighteningly good' in two releases.

More from AGI Musings