The timing is close enough to be the story in its own right: within days of the Trump administration ordering Anthropic to block foreign nationals from accessing its most advanced models, Fable 5 and Mythos 5, Chinese lab Zhipu (operating as Z.ai) released GLM-5.2 as an open-weight model with no usage restrictions at all (GLM-5.2: Built for Long-Horizon Tasks, Z.ai). Whether that was a coordinated response or a coincidence of release schedules, it landed as a pointed contrast, and the benchmark numbers behind it are hard to wave away as propaganda.
The specs
GLM-5.2 is a mixture-of-experts model with 744-753 billion total parameters but only around 40 billion active per token, trained on 28.5 trillion tokens, and released under the MIT license, meaning free download, modification, and commercial use with no restriction (GLM-5.2, Z.ai). It ships with a 1-million-token context window, and Zhipu’s real architectural contribution is a technique it calls “IndexShare,” which reuses the same attention indexer across every four sparse-attention layers, cutting per-token compute at that full 1M-token context by roughly 2.9x compared to a naive implementation. That’s the detail that actually makes a million-token context window commercially viable rather than a benchmark-only number nobody can afford to run in production.
The benchmark numbers
Zhipu built GLM-5.2 specifically to target long-horizon autonomous coding and engineering work rather than chasing general chat benchmarks, and it’s the top-ranked open-source model on the Artificial Analysis composite leaderboard. On FrontierSWE, which evaluates real, open-ended engineering projects rather than isolated coding puzzles, it scores 74.4%, one point behind Anthropic’s Claude Opus 4.8 and slightly ahead of OpenAI’s GPT-5.5. On Code Arena, a blind-comparison coding leaderboard drawing on millions of user votes, Z.ai says it ranks first among globally available models (Zhipu AI’s GLM-5.2 closes in on closed-source leaders in coding marathons, The Decoder).
| Model | Type | FrontierSWE | Notable position |
|---|---|---|---|
| Claude Opus 4.8 | Closed | ~75.4% | Frontier leader |
| GLM-5.2 | Open (MIT) | 74.4% | Top open-weight model |
| GPT-5.5 | Closed | Slightly below GLM-5.2 | — |
What it actually costs
This is the detail that matters most commercially. Across API resellers, GLM-5.2 runs roughly $1.40 per million input tokens and $4.40 per million output tokens, compared to GPT-5.5’s $5/$30 and Claude’s frontier-tier pricing well above that (GLM-5.2, OpenRouter). Blended at a typical 2:1 input-to-output ratio, that works out to roughly a 5.5x cost gap against GPT-5.5, for a model landing within a single benchmark point of the closed frontier on real engineering tasks. An open-weight model that competes with GPT-5.5 on coding benchmarks, at a fraction of the API cost, with no usage restrictions, changes the calculus for a lot of teams currently paying frontier US-lab pricing purely for coding-agent workloads.
Why “open-weight” is the part that actually matters
Benchmark leapfrogging between labs happens every few months and rarely changes anything structural. What’s different about GLM-5.2 is that it’s released with open weights while performing competitively with closed frontier models, a combination the US labs have deliberately avoided offering at this capability tier. That means GLM-5.2 can be self-hosted, fine-tuned, and run without a subscription or API dependency on any single company, which is precisely the profile that makes it attractive both to cost-conscious teams and to anyone uneasy about depending on a model that a government could restrict access to on short notice, exactly what just happened to Fable 5 and Mythos 5 before that ban was lifted (see our coverage of the Fable 5 restoration).
If you want to actually run it
A 750-billion-parameter model is not something you run on a gaming GPU, even with only 40 billion parameters active per token, the full weight set still has to live in memory somewhere. If you’re evaluating GLM-5.2 for local or self-hosted deployment and want the actual VRAM and hardware requirements rather than the headline benchmark numbers, our sister site TensorRigs has a dedicated GLM-5.2 hardware guide covering what it takes to run it at various quantization levels.
The pattern this fits into
GLM-5.2 isn’t an isolated data point. It’s the latest in a sequence, DeepSeek’s releases over the past two years being the most prominent earlier example, of Chinese labs matching or beating closed Western frontier models on specific benchmarks while releasing the weights openly, at a fraction of the price. Each time this happens, the initial Western reaction tends to focus on whether the benchmark numbers can be trusted; the more consistent finding across independent replications has been that they mostly hold up, at least on the specific tasks being measured, even if the models can lag on breadth or on tasks the labs didn’t specifically optimize for.
What makes this particular release land harder is the timing relative to the Fable 5 freeze. A US export-control action aimed at limiting who can access a frontier American model, landing within the same stretch of weeks a frontier Chinese model shipped open-weight and unrestricted to literally anyone, undercuts the practical effectiveness of that kind of control as a strategy. Export controls on model access only constrain the field if the controlled model has no substitute; if an open-weight alternative within striking distance of the frontier is one download away, the restriction mainly affects US and allied users who would have used the restricted model legitimately, while doing comparatively little to limit access for anyone determined to get equivalent capability elsewhere. That’s the strategic tension policymakers are going to keep running into as more labs outside the US reach frontier-adjacent capability, and it’s the same tension running underneath OpenAI’s own restricted rollout of GPT-5.6 that same month, national-security gating only works if the thing being gated is actually scarce.
