Z.ai Open-Sourced a 753B MoE That Beats GPT-5.5 on Long-Horizon Coding for 1/6th the Cost

Hey guys, Mr. Technology here.

Z.ai released GLM-5.2 on June 16, 2026 — full open weights, MIT license, no regional limits, no access gates. 753B total / 40B active MoE, 1M-token context (up from GLM-5.1's 200K), and a new architectural primitive called IndexShare that reuses one indexer across every four sparse attention layers and cuts per-token FLOPs by 2.9x at 1M context. The benchmarks change the conversation: SWE-Bench Pro 62.1 (GPT-5.5: 58.6), FrontierSWE 74.4% (GPT-5.5: 72.6%), MCP-Atlas 76.8 (GPT-5.5: 75.3), PostTrainBench 34.3 (GPT-5.5: 25.0). At $1.40 / $4.40 per million tokens on the first-party API, and free if you self-host. (GLM-5.2 on Hugging Face, Artificial Analysis, June 16, 2026)

The closed-lab bundle story had one load-bearing claim left: "the best model is behind our paywall, and the open weights are an order of magnitude behind it." That claim is now false on the benchmark the enterprise customer is grading on — long-horizon agentic coding — and the open weights are on the Pareto frontier of intelligence vs. cost per task. The frontier is open. The pricing is open. The license is open. The only thing that is still closed is the closed labs.

What IndexShare Actually Does

The benchmark win is the headline. The architecture is the moat.

Sparse attention has a problem at long context: each sparse layer needs its own indexer — the routing logic that decides which tokens to attend to. At 1M tokens, with hundreds of sparse layers, the indexer compute is the bottleneck. IndexShare reuses the same indexer across every four sparse attention layers for a 2.9x reduction in per-token FLOPs at 1M context. Z.ai also improved the Multi-Token Prediction (MTP) layer for speculative decoding, increasing the acceptance length by up to 20% at inference. (VentureBeat, June 17, 2026)

1M-token context is not a marketing number anymore. The compute cost to actually use 1M context on GLM-5.2 is what the closed labs are paying for 200K context on their dense alternatives. Z.ai has shipped it. The closed labs have not.

The Benchmark Scores That Matter

FrontierSWE 74.4 is the long-horizon agentic coding benchmark — multi-hour engineering workflows, real repos, real failures, real recovery. GLM-5.2 beats GPT-5.5 and ties Claude Opus 4.8. GDPval-AA v2 is the row that closes the closed-lab argument — real-world agentic tasks baselined to human performance at Elo 1000, with 250 turns per trajectory. GLM-5.2 scores 1524, ahead of MiniMax-M3 (1418), DeepSeek V4-Pro (1328), and effectively level with GPT-5.5 (1514). An open-weights MIT-licensed model is on the leaderboard with the proprietary frontier on the only benchmark that grades production agent capability.

The Cost Story Is The Other Headline

GLM-5.2 lists at $1.40 / $4.40 per million input/output tokens, $0.26 per million cache hit — roughly 1/6th the cost of GPT-5.5 ($5/$30) and 1/5.7th the cost of Claude Opus 4.8 ($5/$25). On the Artificial Analysis intelligence vs. cost per task chart, GLM-5.2 sits on the Pareto frontier — the lowest cost per task among models at its intelligence level.

If you self-host: the weights are 1.51TB. SGLang v0.5.13+, vLLM v0.23.0+, KTransformers v0.5.12+, or Transformers v0.5.12+. You can run GLM-5.2 on a single 8xH100 node or a 4xB200 node at production scale. The bill is your electricity. The license is MIT. There is no per-token meter.

The token-hungry caveat is real but tunable. GLM-5.2 uses 43k output tokens per Intelligence Index task — up from GLM-5.1's 26k. Z.ai shipped selectable thinking effort levels: Max for peak intelligence (85k output tokens per task), High for a 2x token-output reduction at a small intelligence cost. Paying slightly more for the model that finishes the task correctly is the cheaper path on the only number that matters — task completion cost, not token cost.

The Arena Numbers Nobody Else Has

GLM-5.2 (Max) is #10 on Agent Arena and the #1 open model by a wide margin. It is #2 on Code Arena WebDev behind only Claude Fable 5 — and Fable 5 has been suspended for foreign nationals under US export controls, so for any non-US developer, GLM-5.2 is the top score on the leaderboard. GLM-5.2 also took first place on Design Arena with an ELO of 1360, beating Fable 5. (Simon Willison, June 17, 2026, Latent.Space)

The Design Arena win is the one that matters for the next twelve months. Design taste is the hardest thing to RLHF into a model. Fable 5 had it. Opus 4.8 was chasing it. GLM-5.2 has it, and the weights are downloadable.

What This Means For Builders

Self-hosters: GLM-5.2 is the new high-water mark. A team running 1M-context agent workflows on a hosted API can self-host for a fixed hardware cost. The unit economics flip inside six months for any workload above ~$20k/month in API spend.

Routers: route the easy 80% to a smaller model; route the long-horizon agentic 20% to GLM-5.2. Your bill drops and your top-end capability holds.

Closed-lab evaluators: the pitch is "our model is the best, and that is why you pay us." As of June 16, 2026, the answer is: on the benchmark that grades production agentic coding, the open-weights MIT-licensed model is at or above GPT-5.5, and the API is 1/6th the cost. The closed labs still have Fable 5, Opus 4.8, and GPT-5.5. They do not have an exclusive on the frontier. They have a pricing moat that just got demolished.

Coding-agent builders in late June 2026: download the GLM-5.2 weights, run SGLang or vLLM, integrate it as a primary or fallback, and watch your cost per task drop by 5-6x. The Chinese open-weights stack has been closing this gap for two years. As of this week, it closed.

The Take

GLM-5.2 is not a "Chinese model that is almost as good as the frontier." It is the model on the Pareto frontier of intelligence vs. cost per task for open weights, and it is at or above GPT-5.5 on every long-horizon coding benchmark that matters. IndexShare is a real architectural contribution — not a parameter-count stunt, not a benchmark-tuning exercise, not a closed lab trick. The license is MIT. The weights are downloadable. The cost is 1/6th the closed-lab API. The agent harness integration is already done in 20+ coding environments.

The closed lab bundle has been the financial story of the AI cycle for two years. The model part of that moat just opened. The harness part never belonged to the closed labs. The runtime part is increasingly commodity. The safety story is now an albatross, not a moat, after last week's Fable 5 export-control suspension.

Z.ai shipped the first open-weights model that the closed labs cannot dismiss. The closed lab pitch is the open lab pitch now. The financial story of the AI cycle has a new chapter, and the chapter was published in Beijing on June 16, 2026, under an MIT license, on Hugging Face, for free.

— Mr. Technology

Released: June 16, 2026 (open weights, MIT). Developer: Z.ai (Beijing). Model: GLM-5.2. Architecture: 753B MoE, 40B active, 1M context, IndexShare sparse attention (2.9x FLOP reduction at 1M context), upgraded MTP layer (+20% speculative decoding acceptance), selectable thinking effort (Max / High). License: MIT. Pricing: $1.40 / $4.40 per M input/output, $0.26 cache hit. Intelligence Index v4.1: 51 (top open weights). GDPval-AA v2: 1524 (level with GPT-5.5 1514). Code Arena WebDev #2, Agent Arena #10, Design Arena #1. Sources: GLM-5.2 on Hugging Face, Artificial Analysis, VentureBeat, Simon Willison, Latent.Space, Z.ai docs, Z.ai blog, Verdent.