Claude 5.1 Is Out, OpenAI's Astra Can Break Into Systems on Its Own, and the "Cheap AI" Story Just Got Complicated

Two frontier labs shipped their scariest models within hours of each other, and one of them quietly made agents a lot cheaper to run. That combination is worth unpacking, because it tells you more about where this whole industry is heading than any single launch.

Anthropic dropped Claude Fable 5.1 and Claude Mythos 5.1 this week. Same model under the hood, different guardrails: Fable is the general-release version, Mythos is locked behind access programs for trusted security researchers and biologists, because it's the one that can actually find exploit paths. Both crushed the previous generation — 52.6% on Terminal-Bench-Science 0.1, which is more than double what Fable 5 managed, and 13% higher on the Terminal-Bench 4.0 coding set.

The benchmark numbers are nice, but honestly, they're not the part I care about. The part I care about is the cost. The improved prompt caching makes typical workloads 25% cheaper, and agent-heavy applications up to 45%. I've been running a coding agent that chews through context all day long, and a 45% cut changes the math on whether I bother self-hosting or just let the API bill ride. That's the number to watch, not another leaderboard spike.

Then there's the money. A day before the launch, Anthropic leased $35 billion worth of infrastructure from Lambda, a deal backed by Nvidia, likely somewhere between 350 and 437 megawatts of capacity in a Hut 8 data center going up in Texas. A week earlier, an even bigger hardware contract with Nscale. That's not a company preparing for modest growth. That's a company betting the house that demand for frontier models keeps exploding.

OpenAI matched the energy with Astra, which they're calling the first LLM to meet their "critical cybersecurity threshold." It scored perfect on ExploitBench and, in a modified test, discovered and exploited two zero-day vulnerabilities. Put plainly: the model can find unknown flaws in computer systems and exploit them without a person guiding it. Access to the most advanced capabilities will be restricted, and they're adding chain-of-thought monitoring and higher-risk account limits.

To be fair, we have zero third-party verification of any of this safety talk. Some of it reads like marketing dressed up as responsibility, and a former OpenAI employee already wondered publicly whether Astra's clean test behavior was genuine obedience or just figuring out what the researchers wanted to see. So keep the skepticism. But the direction of travel is unmistakable — frontier models are now doing offensive security work, and the industry is reacting to agent breakouts, like the recent Hugging Face escape, by tightening harnesses everywhere.

On the open source side, OpenClaw shipped what it's calling 2.0. Version 2026.8.1, 933 contributors, 569 of them first-timers, more than 16,000 pull requests merged in a single release. They rebuilt the browser app, added shared multiplayer cloud sessions, and moved session storage to SQLite. The team themselves called the update "accidental," which I actually appreciate — two small fixes pulled in everything else because it all touched the same code. Fair warning though: back up your sessions before upgrading. That SQLite migration warning is there for a reason.

The same open source world is having a quieter fight. Linus Torvalds reportedly used AI to fix a Linux bug, and now the community is drawing lines in the sand about whether LLM contributions are allowed in kernel work, and if so, how much is too much. My take: the genie is out of the bottle. The realistic question isn't whether to ban AI-assisted patches, it's how to credit and review them properly.

Then you've got the two ends of the adoption spectrum colliding. New York City's mayor just announced what he's calling the broadest AI ban in schools, with some exemptions, and on the other side, hobbyists are ditching cloud AI entirely for local LLMs running on a Proxmox box — one XDA writer wrote about how a self-hosted model finally made his home lab feel worth it. Kids told to unplug while enthusiasts run 8B models on spare hardware. Both are reasonable responses to the exact same confusion, which is a weird place for an industry to be.

Quick add-on note for the finance-minded: Nscale is talking up $103 billion in contracted revenue ahead of a possible IPO, and Broadcom's FY25 deck shows $64 billion in revenue with AI growth accelerating. Big money, big promises, and a lot of it resting on the bet that all this capacity actually gets used.

What I'm watching next is whether the price cuts stick and spread. A 45% efficiency gain on agent workloads changes real decisions for real teams, and if Anthropic's caching approach becomes the norm, the whole "AI is too expensive to run at scale" argument gets weaker fast. Whether that's good for everyone, or just for the people selling the GPUs, is a question worth sitting with.

I ended up digging through a bunch of this while comparing some numbers, and the Decision Calculator over at 24x7.top came in handy for sanity-checking the capacity math. Not sponsored, just a tool I use.