GLM-5.3 for Coding Agents: What to Know

Published August 13, 2026 · by the AQ team

Z.ai released GLM-5.3 on August 14, 2026 (Beijing time; August 13 in the US). It is an unusual release: the base model is exactly GLM-5.2, and every gain comes from scaled post-training, more reinforcement-learning environments, more diverse long-horizon tasks, and more compute spent on them. Z.ai's claim is direct: the most capable model of the open-weights lineage for coding, with the top scores in that class on Terminal Bench 3.0 and Agents' Last Exam, though the open-weights title only fully comes due when the weights actually ship. One catch matters for anyone who runs open models: the weights are not out yet, and neither is general API access. At launch, GLM-5.3 (post-trained on the 743B base model) is available through Z.ai's GLM Coding Plan and its ZCode environment, plus an initial group of partners whose services route through Z.ai's official endpoint under its usage policies. Z.ai says API access and the open weights follow in stages after safety evaluations, with the blog naming roughly two weeks for the weights.

What post-training alone bought

The published deltas over GLM-5.2 are concentrated on agentic coding, and some are dramatic. On Z.ai's numbers: Terminal Bench 3.0 rose from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9, SWE-Marathon v1.1 from 19.4 to 42.5, Agents' Last Exam (ALE-CLI) from 23.8 to 28.5, and Terminal Bench 2.1 from 81.0 to 88.2. Z.ai attributes all of it to environment scaling: synthesized, verifiable task environments that look like real units of engineering work (multi-day infrastructure tasks with access to clusters, storage, docs, and experiment results), with agent-generated verifiers checked against oracle, no-op, and unsolved-state tests before their rewards are trusted for training.

Read the numbers with the usual skepticism, and with one unusual credit. They are Z.ai's own, and its headline 50% coding improvement comes from Z.ai Code Bench, an internal benchmark with no public leaderboard. But the comparison table is honest enough to show losses: GPT-5.6 Sol leads Terminal Bench 3.0 outright (34.6 vs 28.3) along with Claude Fable 5 (33.7), Kimi K3 edges it on Terminal Bench 2.1 (88.3 vs 88.2) and leads SWE-Marathon (48.1 vs 42.5), and the closed frontier stays clearly ahead on several rows. The open-weights race is a different story: on Terminal Bench 3.0, GLM-5.3's 28.3 laps Kimi K3's 17.4, and its 28.5 on ALE-CLI is the best score published by any model in the open class. The footnotes are worth reading too: Z.ai ran most of these evaluations inside the Claude Code harness, which says something about how model labs themselves now measure coding ability, in a real agent harness, not a chat window.

Token efficiency, not just scores

The more operationally interesting claim is efficiency. On Z.ai Code Bench, GLM-5.3 at max effort reaches 34.5% task completion at roughly 75K output tokens per task, where GLM-5.2 needed 96K tokens to reach 23.4%. At high effort it reports 31.4% at around 50K tokens, which Z.ai positions above Claude Opus 4.8 (29.5% at 120K) while conceding Claude Fable 5 stays ahead at 39.5%. Internal benchmark, vendor's own chart, but the shape of the claim (better results on materially fewer output tokens) is the one that shows up directly on your bill if it holds up in independent testing.

The cyber angle, and why the weights are delayed

The release's most discussed section is what Z.ai calls emergent cyber capability. Scaled post-training on vulnerability-discovery environments moved GLM-5.3 to the top of CyberGym (84.5%, ahead of every closed model in Z.ai's chart) and more than doubled GLM-5.2 on exploitation benchmarks, though the closed frontier remains far ahead there. Z.ai also reports that, working with security teams, the model surfaced 2,436 vulnerabilities across 269 open-source projects (1,097 critical or high severity), now tracked on a public disclosure ledger at cvd.z.ai; the oldest flaw dates to 1981. This is also the stated reason the weights ship two weeks late: additional safety evaluation and hardening first. For teams that only care about coding, the practical takeaway is the timing, not the security research.

Thinking is now always on

GLM-5.3 changes the API contract: thinking can no longer be disabled. The model exposes three effort levels (low, high, max), defaults to max, and Z.ai recommends max for coding. If your setup pinned thinking.type: disabled, it must migrate before switching model IDs. That has two consequences for agent work. First, latency-sensitive interactive use now always pays some reasoning overhead; there is no fast path. Second, harness handling of the reasoning stream matters even more than it did for GLM-5.2, and scores will vary by harness accordingly. Evaluate it inside the harness you actually intend to run, not from someone else's leaderboard row.

Pricing and where to run it, today vs in two weeks

At launch, everything runs through Z.ai's own service. The GLM Coding Plan and ZCode have GLM-5.3 now, metered with a points system (input, cached input, and output priced separately, with a 50% off-peak discount outside weekday 14:00 to 18:00 Beijing hours, and a limited-time quota boost through August 31). General API access rolls out in stages after safety evaluations, and an initial partner group offers GLM-5.3-powered services routed through Z.ai's official endpoint; Z.ai has not yet named the partners publicly. US-hosted serverless providers cannot list it independently until the weights are public. For reference while you wait, those hosts serve GLM 5.2 today at around $1.40 per million input tokens and $4.40 per million output with the full 1M context window. Once the GLM-5.3 weights land (Z.ai's earlier releases in this series shipped under the MIT license, though the 5.3 license is not yet posted), expect US hosting to follow quickly, the way it did for GLM-5.2 in June.

When to pick it, when not to

Pick GLM-5.3 when the work is agentic and terminal-shaped, where its gains are concentrated and its open-weights records sit, when token efficiency at high effort matters to your unit economics, or when you want frontier-adjacent coding with an open-weights trajectory: audit, self-hosting, and jurisdiction options arrive with the weights. Wait or look elsewhere if you need weights on day one (two-week gap), if you need to disable reasoning for latency reasons (no longer possible), if unguided planning quality is your bottleneck (closed frontier models still lead), or if your harness cannot pass reasoning content through multi-turn tool loops cleanly.

Where AQ fits

AQ is the multiplayer coding harness where engineering teams run AI coding agents like Claude Code and Codex together: shared live terminals, a code editor, and app previews, in your own cloud. GLM 5.2, the model GLM-5.3 is post-trained from, is available in AQ through AQ Models right now: pick it from the model menu when launching OpenCode or Pi in a workspace and it runs with no Z.ai account, no provider key, and no separate billing setup. That inference consumes AQ credits, priced at provider cost during early access, with $5 in credits included when you sign up, and AQ adds no markup on your model usage. AQ's catalog picks up new open-weights models as they land on its hosting, and each goes live in the menu after it clears pricing validation and per-harness compatibility checks, so expect GLM-5.3 there shortly after the weights are public and hosted; this page will be updated when it is live.

The two-week gap is actually a useful evaluation window. Open two workspaces on the same repository (each is an isolated git worktree), give GLM 5.2 and your incumbent model the same backlog item, and let the team watch both live sessions in the browser; our guide to reviewing an AI coding session covers what to compare beyond the final diff. When GLM-5.3 lands, you rerun the same comparison and see exactly what the post-training bought on your codebase, not on a benchmark. Sessions run in persistent tmux on your team's VM, so a long run keeps executing after your laptop closes.