GLM-5.2 is the step change for open agents

Nathan Lambert published this on interconnects on 2026-06-22, six days after the weights of Z.ai's GLM-5.2 came out. The release happened in an odd order. Z.ai rolled the model out on Saturday, June 13, to subscribers of its GLM Coding Plan, and published the MIT-licensed weights and the release blog three days later, on June 16. Lambert notes that weekend releases usually mean something went wrong (Llama 4 is his example). His reading here is that z-ai wanted to catch a news moment. A little over a week earlier the US government had imposed an export restriction on Claude Fable 5 that amounted to an effective ban, and Anthropic was being criticized as anti-open-science for silently adding safeguards against AI researchers. Chinese open-weight labs, he says, have taken every easy marketing win like that for a year or two.

The number suggested an incremental update to GLM-5.1. Lambert's first lesson is that minor versions can cross thresholds that change what a model is good for. By this point Moonshot (Kimi) and Z.ai (GLM) had become the two most liked open-weight labs among AI researchers.

The reaction mattered more than the release notes

Lambert skips most of the launch material: the benchmark scores, the SLIME RL framework, the advice to always run the model at Max thinking effort. Benchmarks are "half dead", he says, and the real test is what other people find. Here the community results beat expectations. On Arena's agent leaderboard GLM-5.2 was the only open model competing with the latest OpenAI and Anthropic models, with GLM-5.2 at max effort matching Opus 4.8 with thinking off. Design Arena, a benchmark that designers themselves distrust, ranked it above Claude Fable.

Lambert says only one open release before this had drawn this much attention from the researchers and commentators he respects, and that was DeepSeek R1. He had once called Kimi K2 a "DeepSeek moment", and he says GLM-5.2 goes well beyond it. K2 showed that big jumps in open models could come from any Chinese lab. GLM-5.2 is "more of a one way door".

The model that feels right in a harness

The central claim is that GLM-5.2 is the first open-weight model that "feels right in coding harnesses as a general agent". Lambert compares it to R1 again. R1 showed open labs with far fewer resources could reproduce OpenAI's o1-style reasoning. GLM-5.2 shows they can compete in agentic coding, the capability behind Anthropic's record revenue growth from claude-code. He points out this was not guaranteed, because agents depend on tools, integrated harnesses and very large models, all of which make them expensive to build.

He tested it himself, using GLM-5.2 through Fireworks' API inside Claude Code to write material for his post-training course, and says setup was easy. There was one annoyance. When Claude Code or his repository docs tried to send an image to the text-only model, the Fireworks session broke and he had to clear the context by hand. Otherwise the model "immediately felt right". Others were more enthusiastic still. Z.ai's founder told Elon Musk that "open-weight Fable capabilities will be here sooner than Q1 2027", and Vercel's CEO said he was "almost shocked" at how good it was at coding.

That a Chinese open model now works inside Anthropic's own agent is the pattern the vault keeps recording. telnyx serves GLM-5.2 as an FP8 build next to Kimi models, glm52-amd-mi355x is a vendor's account of quantizing it to MXFP4 and serving it on AMD hardware, and mesh-llm reports it running across two Mac Studios.

Measuring the gap

Lambert had predicted an "explosion in usage" once an open model reached the level of Claude Opus 4.5 inside Claude Code, which was the frontier at the start of 2026. Opus 4.5 shipped on November 24, 2025. GLM-5.2 shipped on June 16, 2026, 204 days or about 6.8 months later. That falls inside the 6-to-9-month lag usually claimed between American closed labs and Chinese open ones. Lambert says the result surprised him, because he had expected rapid US compute growth to widen the gap, and that Claude Fable 5, a model that depends more on scale and so on the newest GPUs, would test that. He leaves the question open. A month later his kimi-k3-open-weights-escalation essay estimated the gap at 3 to 5 months.

Economic and political consequences

The immediate economic effect is pricing pressure on Anthropic from companies spending heavily on tokens. Lambert does not think Anthropic will miss its revenue forecasts because of it, since demand keeps growing. The bigger winners are companies that sell open-model inference or fine-tuning, including Fireworks, Together, Thinking Machines through Tinker, and Prime Intellect. He expects slow diffusion into real workflows, where teams already use different models for planning, coding and subagents. He also calls it a "severe economic dagger" that this diffusion happened while the leading American model was banned. It gave GLM-5.2 time to take the lower-margin work frontier labs are trying to leave behind.

He treats the political consequence as the more important one. GLM-5.2's release date ties it in people's minds to the Claude Fable restriction and to Claude Mythos. The US government had judged Mythos-class capabilities unsafe to release, while Chinese labs kept releasing more capable models to anyone. Lambert says the two trends are not necessarily causal, since nobody knew how GLM-5.2 performed on cyber tasks compared with earlier versions, but they are correlated. That opens a plausible future in which Washington declares a specific Chinese open-weight model unsafe for the public. us-scrutiny-of-chinese-model-use covers the congressional inquiries that followed in the same months.

His closing argument is about which risk is larger. He understands why an openly available Mythos-class model is frightening. But if open models were banned now while closed models got 10 or 100 times better within two years, in the hands of one or two companies, "we will have bigger problems on our hands." arguments-against-open-source-ai and six-months-to-live-for-open-models take up the same concern from other angles.

What came next

Eight weeks later Z.ai announced GLM-5.3, the same base model with much more post-training. Lambert's glm-5-3-how-chinese-labs-keep-stride looks back on GLM-5.2 as a release that lived up to its hype. Researchers he knew were still using it weeks afterwards for its speed and simplicity.