TL;DR: Z.ai shipped GLM-5.3 on August 14, 2026 — same 744B MoE base as GLM-5.2, all gains from post-training — and it now posts the best open-weight score on Terminal-Bench 3.0 at 28.3%. The catch: at launch you can’t buy it per-token on Z.ai’s general API. It’s reachable only through the GLM Coding Plan (Lite $18/month) and the ZCode agent, wired into Cursor or Cline via the coding endpoint. Open weights are promised around August 28, which changes the self-hosting math again.

	GLM-5.3 (Coding Plan)	GLM 5.2 (general API)	Gemini 3.7 Flash	Claude Fable 5
Best for	Long-horizon Cline agent runs on a flat bill	Pay-per-token GLM work	Cheap fast BYOK backend	Frontier reasoning
Price	$18/mo Lite (10K weekly credits)	~$1.40 / ~$4.40 per M tokens	pay-per-token, budget tier	$10 / $50 per M tokens
Terminal-Bench 3.0	28.3% (top open-weight)	4.6%	not a TB 3.0 leader	34.0%
Weights	open ~Aug 28 (pending)	MIT, on Hugging Face	proprietary	proprietary
The catch	No general per-token API yet	Older post-training	weaker on terminal agents	7× the cost for daily loops

Honest take: If you already run Cline all day and your model bill is the pain point, the $18/month Lite plan is the cheapest way to put a near-frontier agentic model behind it right now — take it. If you’re a Cursor user with a working Pro subscription, wait two weeks: the BYOK override still breaks Cursor’s built-in models, and once the open weights land around August 28 you’ll have router and self-host options that don’t touch that setting. Don’t buy an annual plan for a model whose weights go free in a fortnight.

What actually shipped on August 14

Zhipu AI (Z.ai internationally) released GLM-5.3 on August 14, 2026 as a post-training upgrade of the same base model behind GLM-5.2: a 744B-parameter mixture-of-experts transformer with roughly 40B active parameters per token — 256 routed experts plus one always-on shared expert per sparse layer, 8 routed experts firing per token across 75 layers. Launch coverage reports the same 1M-token context window as GLM-5.2 with a 128K output cap. Nothing about the architecture changed; every reported gain came from scaled-up post-training aimed at coding, long-horizon agent tasks, and security analysis.

The headline number is Terminal-Bench 3.0: GLM-5.3 scores 28.3%, up from GLM-5.2’s 4.6% on the same benchmark. That’s the best result of any open-weight model, and Z.ai also reports an 84.5% CyberGym score and a roughly 50% jump on its internal Code Bench.

The benchmark reality check

Two things keep that 28.3% honest.

First, GLM-5.3 is the open-weight leader on Terminal-Bench 3.0, not the overall leader. The August 2026 leaderboard has Claude Opus 5 at 42.7%, GPT-5.6 Sol at 34.6%, and Claude Fable 5 at 34.0% ahead of it, with Grok 4.6 behind at 26.5%. A 14-point gap to Opus 5 is real and you will feel it on the hardest multi-step tasks.

Second, don’t read the 4.6% → 28.3% jump against GLM-5.2’s old Terminal-Bench 2.1 score of 81.0. Terminal-Bench 3.0 is a different, much harder benchmark — built by Harbor, the Laude Institute, Snorkel AI, and Turing, it adds GPU-enabled nodes up to H100s, multi-container topologies, live microservices, and strict if-and-only-if grading. Every model’s number cratered when 3.0 replaced 2.1. The right comparison is GLM-5.2’s 4.6% on the same 3.0 suite, and a six-fold improvement from post-training alone is genuinely unusual. We covered how leaderboard resets distort model marketing in our Terminal-Bench 2.1 leaderboard analysis; the same caution applies here, doubly so because 28.3% is still vendor-reported until more independent runs accumulate.

How you can actually buy it today

Here’s the part most launch coverage buries: at publication, GLM-5.3 is not on Z.ai’s general pay-per-token API. Z.ai is staging the rollout behind its safety review, so the model is reachable only through two doors:

1. GLM Coding Plan — the flat subscription that routes GLM models to coding tools through a dedicated endpoint. Every tier now runs GLM-5.3; existing subscribers got it automatically on launch day.
2. ZCode — Z.ai’s own agentic IDE built around GLM-5.3, which consumes the same plan credits.

Verified pricing as of August 24, 2026: Lite $18/month, Pro $80/month, Max $168/month, with roughly 30% off on annual billing (about $12.60 / $56 / $117.60 per month). A couple of pricing trackers list Pro at $72 and Max at $160, which looks like promo-window residue — treat the checkout page as the only number that counts on the day you buy.

The tiers differ in weekly credit allowances — 10K / 60K / 140K credits, refreshing every 7 days — not in which model you get. Two quirks of the credit system matter for real budgeting:

• Off-peak usage burns credits at half rate, and Z.ai defines peak narrowly: Monday–Friday, 14:00–18:00 UTC+8. If you code in US or EU daytime, almost all of your usage is off-peak. That effectively doubles the Lite tier for most readers of this site.
• The plan is contractually limited to supported coding tools — 20+ of them, including Cline, Claude Code, Cursor, and ZCode. It is not a discounted general-purpose API, and Z.ai can tell the difference from traffic patterns.

If you’d rather not subscribe at all, third-party routers picked GLM-5.3 up at launch — Requesty lists it at $1.40/M input and $4.40/M output, the same rate GLM-5.2 runs on Z.ai’s general endpoint.

The endpoint, the model ID, and the smoke test

The GLM Coding Plan uses a separate endpoint from the general API, and mixing them up is the #1 support complaint carried over from GLM-5.2:

• General API (no GLM-5.3 yet): https://api.z.ai/api/paas/v4
• Coding Plan endpoint (GLM-5.3 lives here): https://api.z.ai/api/coding/paas/v4

Both speak the OpenAI chat-completions protocol. The model ID is lowercase glm-5.3 on the wire (Cline shows it as zai/glm-5.3). Before touching any editor config, confirm your key and endpoint pair with curl:

$ curl -s https://api.z.ai/api/coding/paas/v4/chat/completions \
  -H "Authorization: Bearer $ZAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "glm-5.3", "messages": [{"role": "user", "content": "Reply with exactly: ok"}]}'

A working setup returns a normal OpenAI-style JSON body with "model": "glm-5.3" in the response. The failure mode to recognize: a Coding Plan key pointed at the general paas/v4 URL (or a general-API key pointed at coding/paas/v4) comes back as an auth error that looks exactly like a bad key. If you’re staring at a 401 and you’ve triple-checked the key, you’re on the wrong endpoint — swap the URL before you regenerate anything.

Cline: still the clean path

Cline ships a native Z AI provider, so there’s no manual endpoint work:

1. Open Cline settings → API Provider → select Z AI.
2. Paste your Z.ai API key.
3. Set the region — international uses Z.ai’s global entrypoint; the China option routes via open.bigmodel.cn. The region setting controls both the endpoint and which models appear.
4. Pick the model. A fresh setup with a Coding Plan key defaults to zai/glm-5.3; if your dropdown only shows zai/glm-5.2, Cline has detected a general-API key — GLM-5.3 won’t appear there until Z.ai opens the general endpoint.

That last behavior is the problem-and-fix worth writing down, because nothing in the UI explains it: Cline filters the model list by what your key can reach, so an old pay-per-token key keeps the dropdown pinned at 5.2 with no error message. The fix is not a Cline setting — subscribe to the Coding Plan (or wait for general API availability), issue a fresh key, and re-enter it. Alternatively, skip the native provider entirely: choose Cline’s OpenAI-compatible provider, set the base URL to https://api.z.ai/api/coding/paas/v4, and type glm-5.3 as the model ID by hand.

Tool calling and streaming work through the native provider the same way they did for GLM-5.2 — long multi-step Cline runs were the workload Z.ai post-trained for, and it’s where the Terminal-Bench delta is most visible. Our GLM 5.2 backend guide covers the thinking-mode toggle and cost patterns that carry over unchanged.

Cursor: the BYOK trap hasn’t moved

Z.ai maintains an official Cursor integration page in its developer docs, and the mechanics are the standard custom-model dance:

1. Cursor Settings → Models → Add Custom Model (OpenAI protocol).
2. Paste your Z.ai key into the OpenAI API key field.
3. Override OpenAI Base URL with https://api.z.ai/api/coding/paas/v4.
4. Add the model name. Z.ai’s docs say to enter it uppercase in Cursor — GLM-5.3 — matching the convention from GLM-5.2.

It works for chat and agent mode. But the long-standing Cursor limitation is still open on the community forum, most recently reproduced with GLM-4.7: once the base-URL override is active, Cursor’s own built-in models stop working. The override is global, not per-model, so you’re choosing between GLM-5.3 and your Cursor Pro allowance until you toggle it back off. Also unchanged: Tab autocomplete always runs on Cursor’s hosted models regardless of BYOK settings, so GLM-5.3 will never power your completions — only chat and agent runs.

That’s why the verdict differs by editor. In Cline, GLM-5.3 slots in beside everything else. In Cursor, it displaces your subscription while active — fine if you’re deliberately replacing Cursor’s models, annoying if you wanted both.

August 28: the open-weights fork in the road

Z.ai paired the launch with a two-week promise: open weights on Hugging Face around August 28, 2026, after what it calls its most extensive safety review to date — the CyberGym 84.5% score is the reason the security hold exists at all.

When the weights land, three things change:

• Routers get cheaper and more plural. Every inference provider that hosts GLM-5.2 today will race to serve 5.3, and per-token prices will converge below Z.ai’s own rates, as they did within days for 5.2.
• Self-hosting becomes legal and practical — at datacenter scale. Same base as GLM-5.2 means the same footprint: roughly 744 GB at FP8, an 8×H200-class node, not a workstation. If you’re weighing the hardware seriously, our sister site has the numbers in its GLM-5.3 local hardware guide, and expect quantized community builds to follow on aifoss.dev’s radar quickly.
• The $18 plan has to compete with free. The Coding Plan’s remaining pitch becomes convenience and off-peak credit economics rather than exclusivity.

Which is exactly why the two-week window matters for your wallet: the Lite monthly plan is a cheap way to evaluate GLM-5.3 on your real repos this week, but locking in an annual Pro or Max subscription days before the exclusivity evaporates is a mistake.

Verdict

For Cline users, GLM-5.3 on the $18 Lite plan is the best flat-rate agent backend you can buy in August 2026 — mostly off-peak usage makes 10K weekly credits go further than the sticker suggests, and 28.3% on Terminal-Bench 3.0 puts it within striking distance of models costing an order of magnitude more per token. Take it over DeepSeek V4-Flash when your work is long multi-step agent runs rather than high-volume cheap completions, and over Gemini 3.7 Flash BYOK when terminal-heavy agentic work matters more than raw speed.

For Cursor users, hold. The global base-URL override still sacrifices your built-in models, and the calculus flips again on August 28. Re-evaluate when the weights are on Hugging Face and the routers have priced it.

FAQ

Can I use GLM-5.3 pay-per-token on Z.ai’s API today? No. At publication it’s exclusive to the GLM Coding Plan and ZCode, with general API access promised later. Third-party routers such as Requesty list it at $1.40/$4.40 per M tokens if you need per-token billing now.

Does the $18 Coding Plan key work in any app? No — it’s contractually limited to Z.ai’s supported coding tools list (20+ tools including Cline, Cursor, Claude Code, and ZCode). It is not a general-purpose API discount.

Why doesn’t glm-5.3 show up in my Cline model list? Your key is a general-API key, and GLM-5.3 isn’t on the general endpoint yet. Use a Coding Plan key, or configure Cline’s OpenAI-compatible provider manually against https://api.z.ai/api/coding/paas/v4.

Is GLM-5.3 the best model on Terminal-Bench 3.0? It’s the best open-weight model at 28.3%. Overall, Claude Opus 5 (42.7%), GPT-5.6 Sol (34.6%), and Claude Fable 5 (34.0%) score higher.

When can I run GLM-5.3 locally? Z.ai says open weights arrive on Hugging Face around August 28, 2026. Budget for datacenter-class hardware at full precision — the FP8 footprint is roughly 744 GB — or wait for community quantizations.

Sources

• GLM-5.3 overview — Z.AI developer documentation
• Cursor integration — Z.AI developer documentation
• Z AI provider configuration — Cline docs
• Z.ai debuts GLM-5.3 with long-horizon coding, cybersecurity upgrades — SiliconANGLE
• Z.ai launches GLM-5.3 with frontier coding — Unite.AI
• Terminal-Bench 3.0 leaderboard, August 2026 — BenchLM.ai
• GLM Coding Plan pricing: Lite $18, Pro, Max — AI Pricing Guru
• GLM Coding Plan guide: weekly credits and off-peak rates — codingplan.org
• Cursor models fail with BYOK overridden base URL — Cursor Community Forum
• Self-hosting GLM-5.3: the open-weights drop — Apidog
• Z.ai glm-5.3 API pricing — Requesty

Last verified August 24, 2026. GLM-5.3 access, pricing, and the open-weights timeline are changing week to week; check z.ai and the checkout page before subscribing.