GLM 5.3 Launches With Big Coding Gains, No API Yet

Z.ai shipped GLM 5.3 on August 14, and its Terminal Bench 3.0 score jumped from GLM 5.2's 4.6 to 28.3 — the kind of coding-benchmark leap that usually comes with a new model, not a post-training refresh.

Same base, new training

GLM 5.3 is not a new model family. Z.ai's release note says the model "uses the same base model as GLM-5.2," and "every gain comes from post-training" — roughly a month of extra reinforcement learning on synthesized long-horizon environments, some representing days of work for an experienced engineer. The tuning targets agentic coding specifically: terminal tasks, repo-level work, and multi-step tool use.

The numbers that matter

On Terminal Bench 3.0, GLM 5.3 scores 28.3 against GLM 5.2's 4.6 — Z.ai calls this "open-source SOTA," though GPT-5.6 Sol still leads the full field at 34.6. On the older, saturated Terminal Bench 2.1, every model tested sits between 81 and 89, and the 5.2-to-5.3 jump is worth just 7 points with no ranking change, a sign the newer benchmark has more room left to show real movement.

Z.ai also claims GLM 5.3 hits 31.4% on its own in-house Code Bench at high reasoning effort using about 50K output tokens per task, against Claude Opus 4.8 at 29.5% using 120K tokens. That comparison comes from Z.ai's own private benchmark, run by Z.ai rather than by Anthropic, so it's best read as a vendor claim about token efficiency rather than an independently verified result.

What's not live yet

The standalone API is still marked "coming soon" with no date, so every first-party route today runs through the GLM Coding Plan, ZCode, or integrations like Claude Code, Cline, and OpenCode. Open weights are promised in about two weeks, once "safety evaluation and hardening are complete" — the HuggingFace repo for GLM-5.3 already exists but returns a 401 rather than a 404, suggesting the release is scheduled rather than merely planned. Z.ai's own pricing table still stops at GLM 5.2 ($1.40 per 1M input tokens, $4.40 per 1M output), with no GLM 5.3 row yet.

One breaking change to plan for

Anyone integrating GLM 5.3 needs a code change if they ever disabled thinking mode: thinking.type: "disabled" has been removed, and Z.ai says the request will simply fail. The new contract accepts only enabled, paired with a reasoning_effort of low, high, or max — Z.ai recommends max for coding work and suggests migrating with low first before raising it.