#glm-5-3

In Anthropic's tests the open GLM-5.3 model succeeded at 50 of 410 exploit attempts, close to Claude Mythos Preview at 56. What the report shows and omits.

Active parameters, not total parameters, predict per-token cost and throughput. What the activation count tells you, and the four things it cannot tell you.

Z.ai published GLM-5.3's 753B weights on August 28 under a bespoke licence, two days after shipping GLM-5.3-Flash under plain MIT. The clause that differs.

In the 2025-26 record we checked, two labs gave a public reason before holding back open weights. What a stated delay tells a buyer, and what it does not.

Rung names, counts, defaults and whether reasoning can be switched off, across 18 coding-CLI and model-API surfaces, gaps and unresolved cells included.

A dated census of 14 OpenRouter stealth listings since 2025: seven officially revealed, four reported only, two unresolved, and Ox Alpha still open.

GLM-5.3's direct token-metered API went live on August 18. Thinking can no longer be disabled, and the effort tiers are low, high and max only.

Three Chinese labs shipped major releases and their licence stories diverge. A buyer's scoreboard on weights, pricing schedules, and benchmark claims.

A vendor-published disclosure ledger reports thousands of findings across hundreds of open-source projects. What maintainers and security teams should check.

GLM-5.3 reuses GLM-5.2's base model, so Z.ai says every gain is post-training. Terminal-Bench 3.0 jumped 4.6 to 28.3, and weights are promised later.