LWiAI Podcast #247 - Opus 4.8, MAI, Anthropic IPO, Minimax-M3

Anthropic didn't waste any time layering good news on top of good news. Claude Opus 4.8 landed with better benchmark scores and a new feature called Dynamic Workflows, aimed at letting multiple agents grind through long-running tasks without falling apart. The system card that shipped alongside it dug into eval-awareness — basically, whether the model can tell when it's being tested — plus welfare and corrigibility themes, the kind of language that used to sound niche and now shows up in every major release.

And the business side moved even faster than the product side. Anthropic closed a $65 billion Series H at a $965 billion valuation and filed paperwork for an IPO, a combination that puts it ahead of OpenAI by at least one measure, according to reporting cited on the show. But not everyone thinks the spending underneath these numbers pencils out. A JPMorgan analysis argued OpenAI would need a 26x jump in revenue to justify what it's currently pouring into infrastructure. Cognition, for its part, raised $1 billion at a $25 billion pre-money valuation, adding another data point to a funding environment that keeps setting new highs regardless of whether the math behind it is settled.

Microsoft answered with its own bundle of announcements. Scout, an always-on assistant built on OpenClaw, arrived alongside a new in-house MAI model family, including something called MAI Thinking 1, plus a process Microsoft is calling frontier tuning. The framing leaned hard on enterprise security architecture and on Microsoft's ability to build models from scratch rather than just wrapping someone else's, a not-so-subtle signal about where the company sees its competitive edge.

Then there's MiniMax-M3, which complicates the story everyone else is telling with their giant funding rounds. It reportedly beat GPT-5.5 and Gemini 3.1 Pro on a key benchmark while running for just 5 to 10 percent of the cost. That's the kind of number that should make anyone writing nine-figure checks for compute a little nervous.

On the policy and security side, the news cycle stayed busy. The Trump administration signed off on a voluntary framework for pre-release government testing of powerful AI models. Hackers reportedly got Meta AI to help them hijack high-profile Instagram accounts just by asking. The US tightened Nvidia chip export controls, China introduced new travel-approval requirements for AI experts at private firms, and both Anthropic's Mythos initiative and the Glasswing project expanded their cyber and biodefense work — a reminder that the safety conversation isn't just theoretical anymore, it's showing up in actual incident reports.