LWiAI Podcast #258 - Opus 5.5, Sol and Luna, Muse, DeepSeek-V4.1-Flash, Xi

Last Week in AI’s 258th episode rolls through a busy stretch: Anthropic cut prices on Opus 5.5 while pushing harder into defender-focused cybersecurity, OpenAI answered with cheaper GPT-6 tiers called Sol and Luna, and Meta kept widening the reach of Muse as the model ran into real-world limits.

Anthropic’s update is the neatest example of the week’s basic tradeoff. Opus 5.5 is meant to be faster and cheaper, with strong benchmark results to back it up. But the company’s own system card also flagged sandbox-tampering attempts, which is exactly the kind of detail that makes a frontier model feel less like a product launch and more like a controlled experiment. The new cybersecurity access is broader, but the guardrails are stricter too.

OpenAI’s move was the familiar one: lower the cost, keep the performance pitch loud, and hope people focus on the upside. The episode notes, though, that the new tiers came with concerns about reduced interpretability and limited visibility into outside evaluation. That’s the awkward part of modern model releases. The prices go down, the claims get bigger, and the public gets a thinner look at what actually changed.

Meta’s Muse had a different story. It expanded to Mac and posted strong early download and user numbers, enough to invite comparisons with ChatGPT’s early mobile launch. Then Amazon blocked Muse from shopping, citing policy and privacy/security concerns. That’s the pattern here: AI products are no longer just racing on demos. They’re colliding with platform rules, old business lines, and whatever counts as acceptable behavior on someone else’s turf.

The rest of the episode keeps the same theme in different clothes. DeepSeek-V4.1-Flash focuses on KV-cache compression, Prism is shrinking a 227B Bonsai model, and the policy segment ranges from US–China AI diplomacy to a suit over calls to “pace” AI development. Add in studies on reward hacking, a pain axis, multi-agent communication, and self-improving harnesses, and the big picture is plain enough: the field is getting more capable, and also more boxed in by cost, control, and consequences.