Claude Mythos Is Big Tech-Only, CaveMan Cuts Token Spend Threefold

Anthropic has released Claude Mythos — a model built to hunt for vulnerabilities in open-source infrastructure. On the SWE-bench Pro benchmark it scores 77.8%, compared to 53.4% for Opus. But the model won't see a public release: Anthropic justifies keeping it closed by arguing that Mythos is simply too good at finding exploits, even without specialized knowledge.

Instead of a public launch, Anthropic has formed a consortium called Project Glowswing. Its members include AWS, Google, Microsoft, Nvidia, Cisco, and Apple. Companies in the consortium get access to test their own systems and write patches. Anthropic has already surfaced dozens of real bugs in Linux and FFmpeg.

The security-through-openness paradox

The situation echoes 2019, when OpenAI announced GPT-2 and said it wouldn't publish the full version — too dangerous. They published it later anyway, and the decision looks laughable in hindsight. In a few years, the Mythos story will likely repeat itself: technological progress outruns access restrictions.

CaveMan: three times more efficient

Token savings showed up in CaveMan, a skill for Claude Code that forces the agent to respond as tersely as possible — no articles, no intros, no polite preambles. The project has racked up 16,000+ stars on GitHub.

According to tests from the repository: a response that used to take 69 tokens now takes 19 — a threefold savings. The mechanism is straightforward: most tokens in an agent's response are polite filler, not substance. Strip the filler and you get the same output for less money. The author added several intensity levels, including a mode in Chinese, where logographic script is more token-compact than the Latin alphabet.

Practice versus theory

A Google engineer ran into the opposite problem: he used AI agents to build dev tools for SQLite with essentially no oversight. The result: the codebase became unmaintainable. He threw out three months of work and rewrote everything from scratch in pair-programming mode with the agent — with active control at every step.

For Java developers, State of Java — the annual survey from JPoint and Joker — is now open. The survey covers Java versions, frameworks, development tools, deployment strategies, and AI development tools (new this year). Last year, Kirill Tolkachev presented the results in a talk — including cases where developers picked technologies that don't actually exist.

Also worth noting: a tool from Milla Jovovich for long-term agent memory. The agent retains context from past conversations and doesn't require repeated explanations. It scored 100% on the SWE benchmark. What's interesting isn't the tool itself, but the precedent: where celebrities used to launch energy drinks and chips, now they launch open-source repos with critical benchmarks.