GLM-5.3: Coding and Cyber Upgrade on the 753B Base, No Retraining

TL;DR

Last updated August 21, 2026. Weights status and benchmarks get updated as they land.

GLM-5.3 is Z.ai's (Zhipu AI's) upgrade of GLM-5.2, released August 14, 2026. Same 753B mixture-of-experts base, no new pretraining; scaled post-training moves Terminal-Bench 3.0 from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9, and CyberGym from 77.2% to 84.5%. Independently, Artificial Analysis scores it 60 on the Intelligence Index, rank 8 of 182, tying Kimi K3 at about one-fifth the price. API pricing is unchanged at $1.40/M input and $4.40/M output with a 1M-token context. Weights were gated at launch.

What Is GLM-5.3?

GLM-5.3 is the model Z.ai shipped on August 14, 2026 under the tagline "Built to Code. Ready for Cyber Defense." It is not a new base model. Z.ai took the 753B-total mixture-of-experts network behind GLM-5.2 and ran a much larger post-training program against agentic coding, long-horizon tool use, and vulnerability research, then shipped the result as a new model id, glm-5.3, at the same price. Context stays 1M tokens with 128K max output. Reasoning is always on and cannot be disabled; effort levels are low, high, and max, with max the default.

The release-week storyline had two threads. First, the post-training claim itself: a 6x jump on Terminal-Bench 3.0 without touching pretraining is a statement about where capability gains currently come from. Second, the cyber positioning: Z.ai says it pointed the model at 269 open-source projects and found 2,436 vulnerabilities, and it stood up a public coordinated-disclosure ledger at cvd.z.ai. The launch hit the Hacker News front page at 1,170 points ("GLM-5.3: Frontier coding with emergent cyber capabilities").

Morph does not serve GLM-5.3 yet because the weights are gated. Morph serves GLM-5.2 today as morph-glm52-744b at $0.8415/M input and $3.1365/M output with the full 1M context; when the GLM-5.3 weights land, the same serving path applies. See Morph Open Source Models.

GLM-5.3 vs GLM-5.2: What Actually Changed

The honest framing: same body, new training. Z.ai did not retrain the 753B base; it scaled post-training. That makes the deltas below unusually clean as a measurement of what post-training alone buys on a fixed base.

Note the Terminal-Bench version: 3.0 is a harder revision than the 2.1 where GLM-5.2 scored 81.0, which is why GLM-5.2 shows 4.6 here. The 4.6 → 28.3 delta is the meaningful number, not a cross-version comparison. If you are choosing between the two models today: the API costs the same, so on Z.ai there is no reason to prefer 5.2. The reason 5.2 still matters is that its weights are out (MIT) and cheaper hosts serve it; Morph runs it at $0.8415/$3.1365.

Benchmarks: Vendor Numbers and the Independent Check

Nearly all launch benchmarks are Z.ai's own. The one substantial independent datapoint so far is Artificial Analysis: Intelligence Index 60, rank 8 of 182 tracked models, measured output speed 93.2 tokens per second. That ties Kimi K3 (60) and clears GLM-5.2 (53). No SWE-bench Verified figure has been published for GLM-5.3 by anyone; if you see one quoted, it is invented.

On Z.ai's internal Code Bench the claim is efficiency, not supremacy: 31.4% at roughly 50K output tokens per task, which Z.ai charts ahead of Claude Opus 4.8 (29.5% at 120K tokens) and behind Claude Fable 5 (39.5% at max effort). Beating Opus 4.8 at less than half the tokens is the pitch; staying a tier below Fable 5 is the fine print. Treat the whole chart as vendor-internal until someone replicates it.

The Cyber-Defense Story

The differentiated part of the release is not the coding delta, it is the vulnerability-research program built around the model. Z.ai reports CyberGym at 84.5%, which it says edges Claude Mythos 5 (83.8%) and GPT-5.6 Sol (83.6%), and ExploitBench at 54.4%, more than doubling GLM-5.2's 24.4% while still trailing the closed frontier (Mythos 5 posts 78.0%). In Z.ai's ExploitGym harness, GLM-5.3 solved 105 tasks in 2 hours and 130 in 6, versus 29 and 39 for GLM-5.2.

Applied, Z.ai says the model surfaced 2,436 vulnerabilities across 269 open-source projects, 1,097 rated critical or high, the oldest introduced in 1981. Disclosure runs through a public ledger at cvd.z.ai, which listed 53 public CVEs at launch with 2,383 still embargoed. This program is also the stated reason the weights shipped late: Z.ai held them back for safety hardening rather than releasing day-one alongside the API.

GLM-5.3 API: Pricing and Access

Today there is exactly one upstream: Z.ai. The model id is glm-5.3 at $1.40/M input, $0.26/M cached input, and $4.40/M output with the 1M context, the same rates as GLM-5.2. It is bundled into the GLM Coding Plan subscriptions ($18/month Lite, $72/month Pro, $160/month Max) and works in Claude Code, OpenCode, and ZCode through Z.ai's Anthropic-format coding endpoint. OpenRouter lists it as z-ai/glm-5.3 (added August 18, 2026) with Z.ai as the single provider; Fireworks, Together, DeepInfra, and Baseten cannot serve it until the weights open.

If your workload runs on open weights now, GLM-5.2 remains the served option: same 753B base, weights on Hugging Face under MIT, and Morph runs it on codegen-tuned kernels behind the standard OpenAI-compatible endpoint:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.morphllm.com/v1",
    api_key="YOUR_MORPH_API_KEY",
)

resp = client.chat.completions.create(
    model="morph-glm52-744b",
    messages=[
        {"role": "user", "content": "Refactor this module and explain the change."},
    ],
)
print(resp.choices[0].message.content)

Where Are the Weights?

Gated. The Hugging Face repo zai-org/GLM-5.3 exists but returns a 401 to unauthorized requests, and Z.ai's stated plan at launch was to release the weights roughly two weeks after August 14, once the cyber capabilities went through safety hardening. GLM-5.2 shipped under MIT with no regional restrictions, and the community expectation is that 5.3 follows, but the license is unconfirmed until the repo actually opens. Until then, "open weights" is announced intent, not current fact, and no third-party host, Morph included, can serve it.

When the weights land, the self-hosting math is the same as GLM-5.2's, because the base is the same: roughly 1.5TB BF16, 8x H200/B200-class GPUs for a single node, FP8 around 744-890GB. This page gets updated when the drop happens.

GLM-5.3 vs Kimi K3

The clean comparison, because it is independent: Artificial Analysis scores both GLM-5.3 and Kimi K3 at 60 on the Intelligence Index. The separation is elsewhere. GLM-5.3 costs $1.40/$4.40 per M tokens against K3's $3/$15, roughly one-fifth the output rate. K3 counters with native vision and with weights that are already public (released July 27, 2026 under Modified MIT, servable by any host), while GLM-5.3's are still gated. Both models run always-on reasoning and both are verbose, so per-task cost depends on effort caps more than sticker price.

For the K2/K3 lineage in depth, see Kimi K3 and the Kimi K3 API guide. For the GLM lineage, GLM-5.2 and the GLM-5.2 API guide.

GLM-5.3: Pros and Cons

When to Use GLM-5.3 (and When Not)

Frequently Asked Questions

What is GLM-5.3?

Z.ai's August 14, 2026 upgrade of GLM-5.2: the same 753B MoE base improved entirely through scaled post-training, aimed at agentic coding and cyber defense. 1M context, 128K max output, always-on reasoning, $1.40/$4.40 per M tokens.

Is GLM-5.3 a new model or a fine-tune?

Neither a new base nor a light fine-tune: same 753B network, but a much larger post-training program. The deltas (Terminal-Bench 3.0 4.6 to 28.3) are what post-training alone bought on a fixed base.

How much does the GLM-5.3 API cost?

$1.40/M input, $0.26/M cached input, $4.40/M output on Z.ai, identical to GLM-5.2, with the 1M context included. Also bundled in the GLM Coding Plan from $18/month.

When do GLM-5.3 weights come out?

Z.ai said roughly two weeks after the August 14 launch, pending safety hardening. The Hugging Face repo is up but gated. License unconfirmed until it opens; GLM-5.2's was MIT.

GLM-5.3 or Kimi K3?

They tie at 60 on the independent AA Intelligence Index. GLM-5.3 is about one-fifth the price; K3 has native vision and open weights today. If you need an open-weight model served on tuned kernels right now, the practical pair is GLM-5.2 and Kimi K3 on Morph.