Z.ai releases GLM-5.3, a frontier coding model with a cyber capability jump that outpaced its training plan

Z.ai has released GLM-5.3, a new flagship version of its GLM model line focused on complex software engineering and long-horizon agent tasks, according to the company's own documentation. The model is live now for GLM Coding Plan subscribers, with API access "coming soon" and open weights to follow after further safety work.

What's new

Z.ai's documentation describes GLM-5.3 as "Z.ai's latest flagship model, delivering major advances in complex software engineering and agent tasks." Unlike prior version bumps, the gains come entirely from post-training rather than a bigger or retrained base model: GLM-5.3 runs on the same underlying architecture as GLM-5.2, with Z.ai attributing the jump to "a more extensive post-training process."

The coding numbers are large. Z.ai says GLM-5.3 "improves by 50% over GLM-5.2 on Z.ai Code Bench," and on more established third-party benchmarks the jumps are steep: "Terminal-Bench 3.0 has increased from 4.6 to 28.3, DeepSWE v1.1 has risen from 46.2 to 66.9, and Agents' Last Exam has improved from 23.8 to 28.5." The company describes these as "open-source SOTA results on benchmarks including Terminal-Bench 3.0 and Agents' Last Exam."

The more notable finding is on the security side. Z.ai's documentation states that "GLM-5.3 reaches SOTA performance on CyberGym for vulnerability discovery" and "doubles GLM-5.2's performance on exploit benchmarks," scoring "84.5% on CyberGym" while "ExploitBench increased from 24.4% to 54.4%." Independent coverage of the release has characterized this cyber capability gain as having grown faster than Z.ai anticipated going into training — a jump large enough that the company is holding back public release of the model's weights, which Z.ai says will follow "in about two weeks" after additional safety evaluation and hardening.

Availability is staged: GLM-5.3 is "available to all GLM Coding Plan users" today through Z.ai's ZCode agent and AutoClaw products, with direct API access still pending and open weights withheld pending the safety review.

Context

GLM-5.3 arrives roughly two months after GLM-5.2 (June 16), which introduced 1M-token lossless context. Z.ai has kept up a rapid release cadence through 2026 — GLM-5, GLM-5-Turbo, GLM-5V-Turbo, and GLM-5.1 all shipped within the first four months of the year — positioning coding and long-horizon agent capability as its primary competitive axis against both Western frontier labs and other Chinese open-weight players like DeepSeek and Alibaba's Qwen. Z.ai's coding models are already integrated elsewhere in the ecosystem: Mistral opened its platform to GLM-5.2 as a third-party model option earlier this month.

Why it matters

The coding benchmark gains alone would make GLM-5.3 a notable open-weight release — a 50% jump on Z.ai's internal coding benchmark and a sixfold increase on Terminal-Bench 3.0 (4.6 to 28.3) is a large single-version jump by any standard, and Z.ai says the model's programming and agent capability now compares to Claude Fable 5. But the more consequential detail is the cybersecurity finding: a near-doubling of exploit-benchmark performance emerging as a side effect of post-training aimed at general coding and agentic skill, not a targeted cyber capability effort, is exactly the kind of dual-use capability jump frontier labs elsewhere (including OpenAI, with its recently reported Astra model) have flagged as a reason to slow releases and add safeguards. Z.ai delaying its own weight release for safety hardening — rather than shipping everything simultaneously, as it has with prior GLM versions — suggests open-weight labs are starting to adopt the same staged-release caution that closed-model labs have used for capability-threshold models, even as commercial access through the Coding Plan ships immediately.