I genuinely think 2026 is the year open source models finally win, and GLM-5 is the best proof I’ve seen so far. In this video, I break down what just dropped: a 744B parameter open-weights “monster” that hits ~50 on the Intelligent Index (the first open coding model to do it) and matches Claude Opus 4.5 on a bunch of real tasks—at a fraction of the cost. The stat I care about even more than size is hallucination behavior: according to the reports I reference, GLM-5 reduces hallucinations massively and actually abstains when it doesn’t know. As a developer, I’ve been begging for that. I walk through benchmarks (including agentic/long-horizon tests like VendingBench and web-browsing task completion) and then I show how I tested it in a real workflow: I used it inside a coding agent to build a comments feature on my course platform. It basically one-shotted the schema and implementation without me spoon-feeding file context, and it didn’t fall apart the way I expected a brand-new open model to. I also get practical about running it: the full BF16 weights are ~1.5TB, so you’re not spinning this up on a MacBook. The point is you can still use it through hosted providers or tools, and when you compare token costs (think ~$0.80/M vs Opus-class ~$15/M), it’s the difference between agentic coding being affordable… or not.