GLM-5.2 takes the open-weights crown — and the gap to GPT-class is now <5 points

What happened

Artificial Analysis published its latest Intelligence Index update this week, and Z.ai's GLM-5.2 now sits at the top of the open-weights leaderboard with a composite score of 68 — narrowly ahead of DeepSeek-V3.2 (66) and Alibaba's Qwen3-Max (65). The Intelligence Index is a weighted average across seven evals: MMLU-Pro, GPQA Diamond, Humanity's Last Exam, LiveCodeBench, SciCode, AIME, and MATH-500. It's the closest thing the industry has to a single number for cross-model comparison, and it's where most serious infra teams check first when picking a base model for fine-tuning.

The headline isn't just that GLM-5.2 won — it's that the top three open-weights models are now all Chinese, and the closest US-licensed open model trails the leader by more than eight points on the same composite. Meta's Llama 4 derivatives, AI2's OLMo, and the various Mistral releases are all clustered well below the frontier the Chinese labs are setting. The previous open-weights leader, DeepSeek-V3.2, held the spot for roughly six weeks. GLM-5.1 launched in March; GLM-5.2 in June. That's a faster release cadence than any frontier US lab has sustained on permissively-licensed weights.

The weights are MIT-licensed and already mirrored on Hugging Face. Inference is live on Together, Fireworks, and DeepInfra at roughly $0.40 input / $1.60 output per million tokens — about 70% cheaper than Claude Sonnet 4.5 at comparable benchmark scores.

Why it matters

The interesting part isn't the score. It's the gap to closed frontier models, which has collapsed faster than almost anyone predicted twelve months ago. GLM-5.2 sits roughly 4 points below GPT-5.1 and 3 points below Claude Opus 4.7 on the same Intelligence Index — a margin smaller than the typical eval noise floor. A year ago that gap was 15-20 points. Six months ago it was 8-10. We are now genuinely inside the regime where 'open is good enough for most production workloads' is a defensible engineering claim, not a hopeful one.

The second-order effect is on inference economics. When the open frontier is 70% cheaper at parity, the closed-model premium has to be justified by something other than raw capability — usually latency, tool-use reliability, or vertical features like Anthropic's computer use or OpenAI's Realtime API. The result is a bifurcating market: closed models compete on agentic features and SLAs, while open weights compete on raw reasoning per dollar. That's a much healthier equilibrium than the 'one lab wins everything' scenario the 2023 discourse assumed.

The community reaction on HN (558 points, top of front page) was unusually substantive. The top comment thread debated whether Artificial Analysis's weighting overweights math benchmarks — a fair critique, since AIME and MATH-500 carry roughly 30% of the composite. Several users posted side-by-side LiveCodeBench-only comparisons where Claude Opus 4.7 still leads by 6+ points. The honest read: GLM-5.2 is the leader on the composite, slightly behind on pure coding, and meaningfully ahead on math-heavy reasoning. If your workload is RAG + summarization + tool-calling, it's effectively tied with the closed frontier. If you're building a coding agent, the closed models still have an edge.

The geopolitical undercurrent is also worth naming directly. Three of the top five open-weights models are now from Chinese labs (Z.ai, DeepSeek, Alibaba) operating under export-controlled compute. Whatever you think the chip controls were supposed to accomplish, slowing the open-weights frontier wasn't it. The labs have adapted — heavy use of MoE architectures, aggressive distillation, and increasingly capable domestic accelerators — and the release cadence has actually accelerated.

What this means for your stack

If you're running self-hosted inference, GLM-5.2 is the new default to benchmark against. The MIT license is genuinely permissive (no acceptable-use clause carve-outs like Llama's), the weights load cleanly in vLLM and SGLang, and the tokenizer is compatible with the GLM-4 family — so existing fine-tuning pipelines mostly just work. The model is 355B total / 32B active MoE, which means it fits on a single 8×H100 node at FP8 and serves at roughly 80 tokens/sec per request.

For teams currently routing through Claude or GPT-4-class APIs for non-agentic workloads — classification, extraction, summarization, RAG synthesis — the cost-quality math now genuinely favors swapping in GLM-5.2 via Together or Fireworks. Expect roughly 65-75% cost reduction at indistinguishable quality on those tasks. The exception is anything tool-use-heavy: GLM-5.2's function-calling reliability is closer to GPT-4o than to Claude 4.7, and agentic loops will see more retries.

For fine-tuning shops, the calculus changed too. GLM-5.2 as a base is now competitive with Llama-4-Maverick for domain adaptation, and the smaller release cadence gap (Z.ai has shipped two major versions in three months) means you're less likely to be stuck on a stale base when the next frontier hop lands.

Looking ahead

The more interesting question is whether the US labs respond with comparable open releases or double down on closed-frontier differentiation. Meta has been quiet since Llama 4; OpenAI's GPT-OSS released in August is solid but not frontier-competitive; Anthropic has shown no signal of releasing open weights at any tier. If the answer is 'closed labs concede the open-weights frontier to Chinese labs,' that's a meaningful strategic shift — and the practical effect is that serious self-hosting infrastructure will increasingly be Chinese-model-shaped by default. Worth thinking about now, before that becomes a procurement question.