Zhipu AI

GLM-5.3

Provider

Zhipu AI

Status

Current

Context

1,000,000 tok

Price

$1.4 / $4.4 /MTok

GLM-5.3 is Zhipu AI (Z.ai)‘s current flagship, released on 14 August 2026 through the Z.ai API and the GLM Coding Plan. It is the most unusual release of the year so far for two reasons. First, Z.ai retrained nothing: GLM-5.3 runs on exactly the same base model as GLM-5.2, and every gain comes from scaled-up post-training — Terminal-Bench 3.0 moved from 4.6 to 28.3 and DeepSWE v1.1 from 46.2 to 66.9 on that basis alone (Z.ai docs). Second, and far more consequential for a lab that built its reputation on MIT-licensed weights, Z.ai withheld the weights at launch — the model’s vulnerability-discovery ability grew faster than the company expected as training scaled, and it delayed the open release by roughly two weeks for safety evaluation and hardening, publishing the weights on 28 August 2026 under its own GLM-5.3 Licence.

The capability claim is independently corroborated. Artificial Analysis scores GLM-5.3 (max) at 60 on its Intelligence Index — one point behind Grok 4.6 and GPT-5.6 Sol Max (61), three points behind Claude Opus 5 (63), and level with Kimi K3 at a fraction of its price — at $1.40/$4.40 per million tokens, unchanged from GLM-5.2. That combination gives it $0.68 per completed index task, the lowest cost per task of any model in the frontier cluster. The honest caveats: it publishes no SWE-bench figure of any kind, it is text-only where its rivals take images, and its open weights carry a custom licence rather than MIT.

Update, 11 September 2026. Artificial Analysis re-based its Intelligence Index on 7 September (v4.3: GPQA Diamond out, Terminal-Bench 4.0 and AutomationBench-AA in), pulling every model’s score down. GLM-5.3 (max) now reads 45, at $2.01 per task — a point ahead of Grok 4.6 and Kimi K3 (44), two under GPT-5.6 Sol (47), and no longer the cheapest per task in the cluster (Grok 4.6 $1.86, Gemini 3.8 Flash $1.24). Index figures elsewhere on this page are the readings at the time they were written, on the old scale, and are kept as dated.

Update, 30 September 2026. Z.ai published GLM-5.3’s open weights on Hugging Face on 28 August 2026 under the GLM-5.3 Licence. At 45 on the re-based index it is the highest-scoring open-weight model Artificial Analysis has measured, a point above Kimi K3 (44).

Quick specs

VIEW GLM-5.3 →

What GLM-5.3 is

GLM-5.3 is a mixture-of-experts model of roughly 750 billion total parameters with about 40 billion active per token — the same architecture, and the same trained base, as GLM-5.2. Z.ai is explicit that this is a post-training release: no new pretraining run, no new foundation, just far more reinforcement learning on long agentic trajectories. That makes the version-on-version deltas unusually clean, because the only variable that changed is the post-training.

It keeps the 1-million-token context and 128K maximum output of its predecessor, and it is text-in, text-out — there is no vision path, which is the clearest capability gap against GPT-5.6 Sol, Claude Fable 5 and Grok 4.6. One migration detail matters if you are upgrading: reasoning is always on. GLM-5.3 does not support thinking.type: "disabled", so applications that used it must switch to enabled with reasoning_effort: "low" (Z.ai docs). Effort levels are low, high and max, defaulting to max.

Benchmark performance

Two things are worth separating: Z.ai’s own launch table, and Artificial Analysis’s independent run.

Independent (Artificial Analysis)

Vendor (Z.ai launch table)

The shape is consistent: GLM-5.3 clears Kimi K3 on the long-horizon terminal work and lands within a few points of the closed frontier, while losing the hardest agentic-coding rows to Fable 5 and Sol. The largest gap is the one that isn’t there at all — Z.ai published no SWE-bench Pro or Verified score, in any harness, having reported 62.1% on SWE-bench Pro for GLM-5.2 two months earlier.

The cyber capability — and why the weights were held

The most-discussed part of this launch isn’t coding. Z.ai reported that GLM-5.3’s vulnerability-discovery ability kept compounding as post-training scaled, past what the company had planned for, with the model reasoning across full exploitation chains rather than finding isolated bugs. The benchmark movement is the steepest on the card: CyberGym 77.2% to 84.5% and ExploitBench 24.4% to 54.4%, with ExploitGym solves going from 29 to 105 tasks inside a two-hour budget (Z.ai docs). On the cyber rows it leads Kimi K3 but stays well behind Fable 5 (181 ExploitGym solves) and Sol (216).

Z.ai says the model has been run against real code with Chinese security teams, reporting 2,436 vulnerabilities across 269 open-source projects, 1,097 of them rated critical or high, tracked through a public coordinated-disclosure ledger — 53 CVEs assigned at launch, with the remainder under embargo. The headline anecdote is a “potentially serious” architectural flaw the model found in Cursor, disclosed privately while a fix is prepared. Treat that one carefully: it is Z.ai’s own account, no CVE has been published for it, and there is no independent confirmation (VentureBeat).

The consequence is the structural story. Z.ai delayed the open-weight release by about two weeks — targeted for late August 2026 — explicitly to complete safety evaluation and hardening first (Axios). It echoes OpenAI’s approval-only GPT-5.6-Cyber four days earlier — two labs in one week putting frontier security capability behind a process rather than a price list, from opposite ends of the open/closed divide. Z.ai published the weights on 28 August 2026 under the GLM-5.3 Licence rather than MIT. It grants MIT-style permissions, except that a company running a model-as-a-service business with more than $10 billion in revenue over any 12 months must pass Z.ai’s security review before commercial use.

Pricing and access

Z.ai lists GLM-5.3 at $1.40 input / $4.40 output per million tokens, with cached input at $0.26 — identical to GLM-5.2 and GLM-5.1, so the capability jump arrives at no price increase (Z.ai pricing). Against the frontier cluster that is roughly a third of Claude Opus 5’s $5/$25 and well under GPT-5.6 Sol’s $5/$30.

The API model ID is glm-5.3, served OpenAI-compatible from https://api.z.ai/api/coding/paas/v4, with an Anthropic-compatible endpoint at /api/anthropic that lets it drop into Claude-shaped agent tooling. The flat-rate GLM Coding Plan includes GLM-5.3 on every tier, from about $12.60/month (Lite, annual billing) to $117.60/month (Max), and Z.ai rolled it out to existing subscribers at launch.

Two access caveats. Hosted use routes through Chinese infrastructure, and Zhipu is on the US Entity List — material for data residency, compliance and government procurement. Self-hosting the open weights, published on 28 August 2026, avoids both.

How GLM-5.3 compares

• vs the frontier cluster — At 60 on the independent index it sits one point behind Grok 4.6 and GPT-5.6 Sol and three behind Claude Opus 5, while beating Grok 4.6 on both of Z.ai’s hardest agentic-coding rows (Terminal-Bench 3.0 28.3 vs 26; DeepSWE 66.9 vs 65.9) and undercutting it on cost per task. What keeps it below them: no SWE-bench figure and no vision.
• vs Kimi K3 — The clearest economic win. The two are level on the independent index (both 60 at max effort on AA’s August re-run), but GLM-5.3 is far cheaper per token ($1.40/$4.40 against $3/$15), cheaper per task ($0.68 vs $0.84), more than twice as fast (84.7 vs 38.2 tokens/sec), and does it at roughly a quarter of K3’s parameter count (~750B against 2.8T) — which also makes it far more practical to self-host. It also clears K3 on Z.ai’s long-horizon rows (Terminal-Bench 3.0 28.3 vs 17.4, AutomationBench 48.2% vs 46.7%). K3’s advantage is native vision; both now have downloadable weights.
• vs GLM-5.2 — Same base model, same price, same 1M context; large gains on long-horizon coding and cyber evals. The trade is the licence: 5.2 is MIT, while 5.3 carries the GLM-5.3 Licence, which adds a security review for the largest model-as-a-service companies.

Known limitations

No SWE-bench figure at all — Pro or Verified, vendor or standardized — which is a real gap for a model marketed on coding, and the main reason we don’t place it higher under a coding-weighted methodology. Text-only: no image input, where Sol, Fable 5, Grok 4.6 and Kimi K3 all take it. Custom licence: the weights, published on 28 August 2026, carry the GLM-5.3 Licence rather than MIT, with a security review for model-as-a-service businesses above $10 billion in revenue; the MIT-licensed GLM-5.3-Flash (26 August) is a smaller, separately trained sibling. China-hosted API and Entity List considerations apply to all current access. Verbose: Artificial Analysis measured 170M output tokens across its eval against a 72M median, so real bills can run above the headline rate. Vendor cross-model comparisons are Z.ai’s own, using competitor numbers pulled from published cards rather than reruns.

Frequently asked questions

What is GLM-5.3?

GLM-5.3 is Zhipu AI (Z.ai)‘s current flagship model, released 14 August 2026. It uses the same ~750B-parameter mixture-of-experts base as GLM-5.2 with substantially scaled-up post-training, keeps a 1-million-token context, and is built for long-horizon coding agents.

Is GLM-5.3 open source?

Open weights, yes; MIT, no. Z.ai withheld GLM-5.3’s weights at launch, citing the model’s faster-than-expected vulnerability-discovery capability, and published them on Hugging Face on 28 August 2026 under the GLM-5.3 Licence, which grants MIT-style permissions but requires a Z.ai security review before commercial use by model-as-a-service businesses with more than $10 billion in revenue over 12 months. On 26 August the company also released GLM-5.3-Flash under MIT — a separately trained 320B-parameter MoE (18B active) that takes text, images and video at a 1M context, previewed anonymously on OpenRouter as “Ox Alpha” from 20 August and priced at $0.15/$0.50 per million tokens hosted. It is a separate, smaller model, not the flagship’s weights.

How much does GLM-5.3 cost?

$1.40 per million input tokens and $4.40 per million output, with cached input at $0.26 — the same rates as GLM-5.2. Artificial Analysis measures $0.68 per completed Intelligence Index task, the lowest in the frontier cluster. The GLM Coding Plan starts at about $12.60/month on annual billing.

Is GLM-5.3 better than Kimi K3?

They are level on capability and far apart on cost. Both scored 60 on the Artificial Analysis Intelligence Index at max effort on AA’s August run (45 and 44 on the re-based v4.3), but GLM-5.3 runs at $1.40/$4.40 against K3’s $3/$15, costs less per completed task ($0.68 vs $0.84), returns tokens more than twice as fast (84.7 vs 38.2 per second), and does it with about a quarter of the parameters. Kimi K3 keeps one real advantage, native vision; both models now have downloadable weights.

How good is GLM-5.3 at coding?

Strong, with a caveat. Z.ai reports Terminal-Bench 3.0 at 28.3 (up from 4.6) and DeepSWE v1.1 at 66.9 (up from 46.2), which clears Kimi K3 on long-horizon terminal work while trailing Claude Fable 5 and GPT-5.6 Sol on the hardest rows. But Z.ai published no SWE-bench Pro or Verified score in any harness, so the most comparable cross-model coding number simply doesn’t exist for this model.

Why did Z.ai delay the GLM-5.3 weights?

Because the model’s offensive-security capability grew faster than the company planned for during post-training — Z.ai’s own card reports ExploitBench more than doubling, from 24.4% to 54.4% (Z.ai docs) — and the company wanted to complete safety evaluation and hardening before releasing weights that cannot be recalled. OpenAI had gated its GPT-5.6-Cyber model behind an approval programme four days earlier, for the same underlying reason. Z.ai published the weights on 28 August 2026.

Last verified 30 September 2026 for the open weights and licence, and 19 August 2026 for the launch figures. The 14 August release date, 1M context, $1.40/$4.40 pricing, model ID and the always-on reasoning change are confirmed against Z.ai’s own documentation and pricing pages; the Intelligence Index score of 60, cost per task and output speed are Artificial Analysis’s independent measurements. The cross-model benchmark table is Z.ai’s launch table, using competitor figures it pulled from published cards rather than reruns — treat it as a ceiling. The Cursor vulnerability is Z.ai’s account with no published CVE and no independent confirmation. The open weights and the GLM-5.3 Licence (28 August 2026) are confirmed against Z.ai’s Hugging Face repository. GLM-5.3-Flash’s MIT release (26 August), specs and Ox Alpha preview are confirmed against its Hugging Face repository and launch coverage.