GLM 5.2
GLM 5.2
Zhipu (Z.ai)'s flagship model, released June 13 2026 at 5:21pm Beijing time — one day after the June 12 US export ban on Anthropic Fable 5. The article's data anchor for the "second China AI moment."
Capability
- Artificial Analysis: most intelligent open-source model on the market. 4th overall (behind ChatGPT 5.5, ahead of Google Gemini).
- Fable 5 is ~17% cleverer on average benchmark tasks (Artificial Analysis composite).
- On private benchmarks:
- ~7 months behind on Weirdml (unusual ML tasks needing careful reasoning).
- ~1 year behind on SimpleBench (common-sense-trap questions).
- On the office-worker exam (Artificial Analysis, released June 19): outperformed ChatGPT 5.5 (2 months old) — GLM 5.2 couldn't have trained for it, so the result is a signal, not a benchmark game.
The four-vs-four-to-ten months lead question
- Naive comparison: GLM 5.2 today ≈ Western model from ~4 months ago (February 2026).
- Havard Tveit Ihle (NDRE): Chinese models score better on public benchmarks (which have published questions) than private ones. Real lead is closer to 8–10 months than 4–6.
- Mechanism: "possibly unwittingly, teach to the test."
- Corroborated by a US government study (May 2026).
Pricing (the buried caveat)
- Per-token pricing: DeepSeek v4 charges $0.87 per 1M output tokens; Anthropic charges $50 for the same on Fable 5. ~57× cheaper per token.
- But Chinese models use many more tokens to reach the same answer:
- Du Zheng (Georgia Tech) et al., updated June 2026: DeepSeek used 23× more tokens than an OpenAI rival to achieve basically the same result.
- Total-cost accounting: on a software-engineering benchmark, GLM 5.2 ended up costing more than systems from Anthropic and OpenAI.
Takeaway: the "Chinese open-source is a fraction of the cost" narrative is often wrong when priced correctly. Per-token is the wrong denominator.
Distribution / access
- Open-weight — can be downloaded and run on local hardware, out of reach of US or Chinese state action.
- API service available but subject to service interruptions and slowdowns during traffic spikes — Chinese compute shortage.
- US regulatory risk: 2 congressional committees investigating American firms using Chinese models.
2026-07-04 Economist citation
America Should Not Imprison Frontier AI (Economist) (Leader): "One recent release, glm 5.2 from z.ai, already matches the best of the last generation of American models." Cited as the specific evidence why a permanent US block on Chinese frontier models is unworkable — Chinese labs "may take longer to catch up with Mythos, since they have fewer chips and American labs are cracking down on distillation… But that buys months, or a year at most."
The 07-04 read reinforces the 06-27 buried-lede finding: GLM 5.2 is capable enough to disprove the permanent block, and the permanent block is what the Hierarchy of Access regime would need to remain durable.
2026-07-06 — Whittemore: the first legitimate "DeepSeek moment" since the original
The Big Ways AI Just Changed (AI Daily Brief) gives GLM 5.2 the label every Chinese release since January 2025 has chased and (in his view) none earned until now. His threshold argument: GLM 5.2 isn't Fable-5-class or even necessarily Opus-4.8-class — but it exceeds the Opus 4.6 / GPT-5.2 level that initiated the agentic era at the turn of 2025/26. That makes it the first open-weight model where the fallback strategy feels "less like compromise and more like genuine competition for the frontier."
Released into the Fable-suspension window, it anchored the diversification wave: custom post-trains on open weights (Cursor's Composer 2.5, built off Kimi), and integrated multi-model architectures using GLM as the workhorse — Harvey + Fireworks paired a GLM worker with an Opus advisor for legal tasks, beating Opus alone at a fraction of the cost (see Model Routing).
warning Hold against China Is Having Another AI Moment (Economist)'s total-cost caveat above: on a software-engineering benchmark GLM 5.2 ended up costing more than Anthropic/OpenAI systems once token overuse is priced in. Whittemore's "genuine competition" frame and the Economist's "per-token is the wrong denominator" frame are both true — the reconciliation is architectural (GLM as routed worker with a frontier advisor) rather than wholesale substitution.
Cross-references
- Zhipu (Z.ai) — the lab
- DeepSeek — the pricing anchor and the 23× token-overuse mechanism
- Anthropic — Fable 5 as the price-and-capability anchor
- Hierarchy of Access — the policy backdrop that opened GLM 5.2's window
- Token Scarcity — the total-cost caveat lands directly in this thread
Sources
- China Is Having Another AI Moment (Economist) — canonical