GLM 5.2 unlocks something SCARIER than Mythos..

GLM 5.2 crossed the line where open models quietly displace closed defaults — and the real headline is the security and routing shift, not the benchmark score.

What GLM 5.2 actually does

The model is reportedly surfacing bugs in complex C++ and Rust code that GPT 5.5 X High misses, and on a Klein harness run it cleaned up dead code where Opus 4.8 shipped with type errors. Against Opus, Not Lobe logged roughly 2x the tokens, faster latency, and ~3x cheaper cost at similar quality. One open-weights report claims GLM 5.2 is the first open model to cross 80% on a specific frontier benchmark covered in the prior day's video.

The security angle Joshua Sax flagged

Sax's framing — GLM 5.2 not Mythos is the real security emergency — is the load-bearing point. Open weights at this capability level let attackers run private long horizon offensive workflows without API logging. When open models cross a threshold, local autonomy and privacy stop being nice-to-haves and become the reason the model gets used.

Local inference is no longer theoretical

Onslaught AI ran a 1-bit GLM 5.2 GGUF on a Mac Studio M3 Ultra with 5× 256 GB memory at ~21.6 tok/s. That puts serious agent workloads on real Apple silicon, not just rented GPUs.

The real winner is the router

JP Schroeder's data: DeepSeek V4 Flash covers ~80% of Claude or Codex tasks at 137x cheaper per task than Fable. A Kimmy paper shows reasoning models overthink and produce worse output anyway. The new stack is cheap model first, escalate only on failure, route by task, and keep it local when privacy or cost matters.

Tooling is catching up

GitHub Copilot now supports bring your own key with Ollama, Foundry, OpenAI-compatible completions, and Anthropic-compatible message endpoints — the plumbing for this routing stack already ships.