GLM-5.3 API released, Terminal-Bench score jumps from 4.6 to 28.3

GLM-5.3 is not a model built from scratch. Its base model is identical to GLM-5.2, released on June 16, with only post-training — the stage where a finished model is retrained on real-world tasks to refine it — added on top. Even so, the results aren't small. On Z.ai's own coding benchmark, the model showed a 50% performance improvement over GLM-5.2, and the company said it achieved the top scores among open-source models on Terminal-Bench 3.0 and Agents' Last Exam (CLI).

The GLM family has been expanding on several fronts in recent weeks. On August 11, Z.ai's coding tool ZCode surpassed 1 million users, and usage limits were reset across all GLM Coding Plan users, with an update to strengthen long-horizon task capability already previewed at the time. The results now seen in GLM-5.3 can be read as that preview taking concrete form. A day later, on August 12, Mistral announced it would bring GLM-5.2 into its European compute alliance platform as a third-party open model — a sign that the GLM lineup is establishing itself not just through Z.ai's own services but as a default option on other platforms as well.

How coding ability improved — by making the environment closer to the real thing

Z.ai described the core of this improvement as "environment scaling." Where existing coding benchmarks tended to be short practice problems, GLM-5.3's training incorporated work-scale tasks that would take a skilled engineer days to complete. For example, in an ML infrastructure task, the model was given the same working environment a real engineer would have — compute clusters, storage, internal documentation, codebases, and experiment results — and asked to diagnose bottlenecks in the training stack, apply optimizations, and deliver a measurable speed improvement while preserving accuracy, all the way through to completion.

Since humans can't manually build this kind of task environment at scale, Z.ai said it built a pipeline that collects real work patterns and automatically converts them into long-running environments with multi-step dependencies and hidden state. A separate judge agent solves each task directly to verify it's actually solvable, and a synthesized verifier — without access to the answer — must pass checks for oracle correctness, no-action baselines, and unsolved states before its signal is trusted as a reward. The company also carried over the SAO (with compression) reinforcement learning strategy introduced in GLM-5.2, so that performance gains hold not just on short tasks but on long-horizon ones as well.

An unexpectedly fast gain in cybersecurity

As the scale of post-training grew, cybersecurity turned out to be an area that improved faster than expected. GLM-5.3 posted the highest score yet recorded on CyberGym, a benchmark for vulnerability discovery. More notably, the gap with GLM-5.2 widened the deeper the exploit chain (reproducing an actual attack) went — the relevant benchmark score more than doubled that of GLM-5.2. Z.ai framed this as being for defensive security purposes, but the fact that vulnerability discovery and actual exploitation ability grew together also means that offensive and defensive capabilities are advancing in tandem.

Benchmark	GLM-5.2	GLM-5.3
Terminal-Bench 3.0	4.6	28.3
DeepSWE v1.1	46.2	66.9
Agents' Last Exam (CLI)	23.8	28.5

How to try it

GLM-5.3 can be accessed in two ways: calling the API directly, or attaching it to a coding agent through the subscription-based GLM Coding Plan.

Via API — Z.ai supports three protocols.

Protocol	Base URL
OpenAI Chat Completion	api.z.ai/api/coding/paas/v4
OpenAI Response	api.z.ai/api/v1
Anthropic Message	api.z.ai/api/anthropic

Users who previously called the API with thinking.type: "disabled" must switch it to enabled, set reasoning_effort to low, and change the model ID to glm-5.3. From GLM-5.3 onward, the option to fully disable reasoning no longer exists. Reasoning intensity is now split into three levels — low, high, and max — with max as the default. For complex coding tasks, using max is recommended.

Via GLM Coding Plan — the model can be connected directly to any preferred coding agent. The new pricing structure has shifted to a points-based metered system, and calls made during off-peak hours, including weekends, consume only 50% of standard points. Anyone who has used a GLM Coding Plan in the past, including expired subscriptions, can currently access the model API only through the OpenAI Chat Completion-compatible protocol.

Who can use it — The official API is open with no waitlist once an API key is issued and a payment method registered, and the GLM Coding Plan can be accessed through individual or team subscriptions. Access via partner model gateways has also been made available.

What you can try — For example, you could hand off an entire task of diagnosing and fixing performance bottlenecks in a legacy codebase, or have an internal security team attach it to an automated vulnerability-scanning pipeline to speed up the discovery stage. The 1M-token context also makes it possible to load an entire large repository at once and have it reviewed in a single pass.

Editor's take

The real thing worth noting in this announcement is the phrase "the base model hasn't changed." Boosting benchmark scores nearly sixfold by pouring additional post-training into the same underlying model — rather than training a new one from scratch — signals that competition in the open-source camp has shifted from a battle over foundation model size to one over how closely the retraining environment mirrors real-world work. The environment-synthesis pipeline Z.ai describes — automatically converting real work patterns into executable tasks and having a judge agent verify them — isn't fundamentally different from what OpenAI or Anthropic have reportedly done behind closed doors. What's different is that this is being applied to an open-weight model, with the resulting scorecard made public.

From experience deploying open-source coding models of similar scale in real workflows, a multi-fold benchmark improvement usually comes down to one of two things: either the ability to see genuinely long-horizon tasks through to completion has actually improved, or the benchmark itself has been redesigned around that model family. With a benchmark that's had a version update, like Terminal-Bench 3.0, the latter possibility should always be kept in mind. Still, gains that show up consistently across benchmarks as different in character as DeepSWE and CyberGym are hard to dismiss as coincidence.

For domestic development teams, the practical takeaway is clear: with pricing held flat against GLM-5.2 while performance rises, any organization already running the GLM family through a coding agent effectively gets a free upgrade just by switching the model ID. That said, because the reasoning-disabled option has been removed entirely, teams running GLM in latency-sensitive, real-time response pipelines should switch to low mode and re-measure response speed before anything else. Security teams drawn to the cybersecurity gains should also weigh the possibility of misuse as an offensive simulation tool, even under a strictly defensive-use premise.

In the coming weeks, it seems likely that platforms bringing in GLM through a gateway, as Mistral has, will move to switch from GLM-5.2 to 5.3. With pricing unchanged, there's reason to switch and little reason not to.