GLM-5.2

What it is. GLM-5.2 (2026-06-16) is Zhipu AI’s previous flagship text model: a 744B-total / 40B-active open-weight model with a 1M-token context window, released under the MIT license. Why it matters. It is the direct predecessor of GLM-5.3 and remains listed on the API pricing page at the same price, making it a live reference point for the post-training gains of its successor. Key characteristics. Text-only input, reasoning in high/max modes, 1M context, 163,840 max output. What a professional should know. It is deprecated and superseded by GLM-5.3 (same base model, improved post-training), so new workloads should prefer GLM-5.3 — but GLM-5.2 remains billable and is worth tracking if you pinned a specific checkpoint.

Architecture and parameters

GLM-5.2 is a 744B-total / 40B-active MoE (~5.4% activation), with open weights in BF16 and FP8. It shares this base with GLM-5.3, which means the entire GLM-5.2→5.3 delta is post-training, not architecture. China AI Hub analysis: that makes the pair a clean controlled comparison for how much post-training alone can move a fixed model. See Mixture-of-Experts.

What the context window actually means

The 1M-token context with a 163,840 output cap sits between GLM-5.3’s 131K and DeepSeek’s 393K on generation headroom. As a long-input analysis tool it matches the frontier norm. See the context-window research.

Pricing implications

$1.40 input / $4.40 output per 1M tokens (cached $0.26) — identical to GLM-5.3. Because the successor costs the same, there is no price reason to stay on GLM-5.2; the pricing page simply has not retired it. See the pricing-changed research.

API, coding, and agent implications

Vendor-reported HLE 40.5 (text-only) / 54.7 (with tools) and SWE-bench Pro 62.1 describe the pre-5.3 capability baseline. The with-tools HLE delta is the same agentic pattern seen across the database. For coding and agent work, GLM-5.3’s claimed post-training gains (including the private Code Bench “50%” improvement) argue for the successor. See GLM agent-oriented AI.

Open weights and license

Open weight, MIT license — described in the official model card as “pure open, no regional limits,” which is notably cleaner than the Apache-2.0-with-verification note on GLM-5.3. Teams that specifically want an MIT-licensed 1M-context GLM checkpoint may prefer 5.2 on licensing grounds despite the deprecation. See licensing explained.

Benchmark interpretation

The three rows are vendor-reported. HLE 40.5 (54.7 with tools) and SWE-bench Pro 62.1 are the GLM-5.2 baseline against which GLM-5.3’s improvements are claimed. The with-tools HLE delta (40.5 → 54.7) mirrors the agentic pattern seen across the database, where tool access raises the ceiling on the hardest knowledge benchmarks. These scores are useful as a historical baseline, not as a current competitive signal.

Context and output in practice

The 1M-token input window places GLM-5.2 at the frontier norm of its release wave, but its 163,840-token output cap is worth noting because it actually exceeds GLM-5.3’s 131,072 ceiling — an artifact of the two models’ separate documentation rather than a capability regression. For long-input analysis (multi-document review, codebase-wide context) the input window is what matters; for generation, both models are long-input tools rather than long-form generators.

What the benchmarks do not prove

Scores are vendor-reported and predate the 5.3 successor; they say nothing about GLM-5.3’s actual gains, which rest on a private benchmark. As a deprecated model, the main utility is historical comparison. See how to read vendor-reported benchmarks.

Suitable and less suitable workloads

Well-suited: text-only long-context analysis where an MIT-licensed (no-verification) open checkpoint is specifically required, and historical/regression baselines against GLM-5.3. Less suited: any new production workload (deprecated, superseded), vision/multimodal input, and cost-optimization (same price as the better successor).

Where this model fits

Workload	Relevance
Long-context analysis	High
Coding	High
Structured API workflows	Moderate
Agent orchestration	High
Local self-hosted deployment	High
GUI automation	No evidence
Video generation	No evidence
Enterprise cloud	Moderate

Relevance judgments are China AI Hub analysis based on documented capabilities, not vendor claims.

China AI Hub analysis

China AI Hub analysis indicates GLM-5.2’s residual value is almost entirely as a licensing option and a historical baseline: its MIT “pure open” label is cleaner than its successor’s Apache-2.0-with-caveat, and its benchmark set is the reference point for GLM-5.3’s claimed gains. For everyone else, the model is deprecated at identical pricing, which makes staying on it a strictly dominated choice — the successor costs the same and improves on the same base.

Market position and outlook

China AI Hub analysis indicates GLM-5.2 illustrates a recurring pattern in the database: deprecation is not retirement. The 2026-09-10 DeepSeek change-log reversal kept DeepSeek-V4-Pro’s API running past its announced cutoff, and GLM-5.2 similarly remains billable on Z.ai’s pricing page at the same $1.40/$4.40 as its successor. The practical consequences are threefold: existing integrations continue to function, the model stays useful as a regression baseline for measuring GLM-5.3’s claimed post-training gains, and its MIT “pure open, no regional limits” label keeps it relevant for teams that specifically want a permissive, no-verification 1M-context GLM checkpoint. For everyone else the model is a strictly dominated choice — the successor costs the same and improves on the same base. See the state of China’s AI models and licensing explained research.

What is uncertain

• It is superseded by GLM-5.3 but remains billable at the same price — a deprecation without retirement.
• Benchmark scores are vendor-reported and predate the GLM-5.3 successor.
• GLM-5.3’s claimed gains rest on the private Z.ai Code Bench and cannot be independently reproduced.

Sources

• Z.ai docs — GLM-5.3 model page
• Z.ai pricing
• Z.ai release notes
• Hugging Face model card — GLM-5.2

Labels used above: Official fact (from primary sources), Vendor-reported claim (benchmark scores published by Zhipu), and China AI Hub analysis (our synthesis, always introduced as such). No third-party benchmark evidence is currently recorded for this model.