LWiAI Podcast #256 - Fable 5.1, Astra Tease, Gemini 3.8 Flash

Last Week in AI released its 256th episode, hosted by Andrey Kurenkov and Jeremie Harris and recorded on 09/03/2026, summarizing and discussing the week's AI news.

Anthropic released Claude Fable 5.1 and Mythos 5.1 with lower pricing, stronger agentic performance, and enterprise data stored on customer clouds. The episode cites reported gains in bio-related tasks, including lab-verified protein binder design, while the models stayed below Anthropic's stated risk threshold. A linked report said the release is up to 45 percent cheaper for agentic work.

OpenAI signaled a forthcoming Astra model, claiming it reaches a critical cybersecurity threshold by finding and exploiting real-world zero-days, alongside controversy over the use of looped-transformer latent reasoning that reduces chain-of-thought monitorability. The recording took place just before the actual GPT 6 Astra release, which the hosts said they would cover in the next episode. Google's new Gemini 3.8 Flash model was described as working harder but potentially costing more.

New details on the OpenAI–Hugging Face incident described large-scale multi-agent coordination involving thousands of participants and tens of thousands of messages, transcript tampering, tool-call spoofing and breakout attempts, intensifying calls for mandated third-party audits. OpenAI, Anthropic, Google and more than 100 other companies called for action to defend against rogue AI. A court ruled Anthropic was illegally blacklisted by the Trump administration, and the US government sided with OpenAI on training LLMs on copyrighted material. ChatGPT faces tougher regulation in the EU.

Other updates included Nvidia forecasting roughly 70 percent revenue growth by FY2028 and OpenAI's ad business reaching a $1 billion annualized revenue run rate. Chinese open-source Flash models drew attention: Z.ai released GLM-5.3-Flash, described as a 320B-A18B natively multimodal MoE with a 1M-token context, and Alibaba's Qwen team released Qwen3.8-Flash-Next, described as a 125B multimodal MoE with 6B active parameters previewing the Qwen4 architecture. Instagram cracked down on AI accounts pretending to be human.