Three stories from this weekend worth paying attention to. Anthropic confirmed a new model tier above Opus after an accidental CMS leak. OpenAI finished pretraining its next frontier model and killed Sora to free up the compute. And Claude's persistent agent threads are rolling out to production. Each one connects to the same underlying questions: what's it going to cost, and how carefully are you managing access.

Anthropic Accidentally Confirms a Model Above Opus

On March 26, a CMS misconfiguration at Anthropic exposed internal draft assets, revealing a new model codenamed Mythos under a new product tier called Capybara. Rather than deny it, Anthropic confirmed the model exists and called it "the most capable we've built to date." The capability framing centered on coding, reasoning, and cybersecurity, with Anthropic describing it as a step-change improvement over Claude Opus 4.6. Early customers are already in pilot. The news sent cybersecurity stocks down 3-7%, with CrowdStrike and Palo Alto Networks both falling as the market priced in the possibility that Mythos handles some of what those tools do.

My Take: A new tier is fine and dandy, but what's it going to cost? How quickly will it eviscerate our tokens? The capability jump sounds real, but for solo founders running automated workflows, the math on token costs matters as much as the benchmark scores. Until we see pricing, the excitement is a little premature.

OpenAI Finishes Training "Spud" and Kills Sora to Get There

Sam Altman told employees in an internal memo that Spud, OpenAI's next flagship model, has finished pretraining and is weeks from release. To free up the compute, OpenAI is shutting down Sora. The product organization was also renamed to "AGI Deployment" as part of the same internal shift. Altman reportedly described Spud as a model that can "really accelerate the economy," and early internal messaging positions it as a meaningful capability leap.

My Take: I think it's a bold move to drop Sora, but I understand it from a business perspective. It probably cost them a significant amount of money for minimal gain, if any. Compute is the constraint right now, and redirecting it toward the next frontier model is the right call. The rename to "AGI Deployment" is telling. That's a company that's decided to stop hedging.

Claude Now Has Persistent Agent Threads Across Mobile and Desktop

Anthropic is rolling out persistent agent threads in Cowork to Max plan users, with Pro following shortly. The same context persists across Claude Desktop and Claude for iOS and Android, with no starting over when you switch devices. Combined with existing computer use capabilities, this turns Claude into a continuous background operator that holds context between sessions, across platforms, and through handoffs.

My Take: I use this feature daily since it's been in preview and it's an absolute game changer for the way I run Chento OS. That said, I would still advise anyone using it to segment access carefully. Claude should not have access to all of your personal files. Set up dedicated folders for work it's allowed to touch, and keep the rest out of reach. The power is real, but so is the surface area.

Three capability jumps in a weekend, and the cost and access questions are trailing close behind. Mythos raises pricing unknowns. Spud reshuffles the compute equation across the whole industry. Persistent threads make Claude more capable of running autonomously across your devices, which is exactly why access scoping matters more now than it did last week. If you're building on AI infrastructure, watch what these releases actually cost before you commit architecture decisions to them.

Read the full post on chento.io