Zhipu AI

Executive Summary

What it is: Zhipu (z.ai) sells GLM models via the credit-transparent GLM Coding Plan (Lite $16.2 to Max $144/mo, 50% off-peak credit deduction, now fully on GLM-5.3), the ZCode desktop IDE, and OpenAI/Anthropic-compatible APIs. GLM-5.3 launched Aug 17 (same base as 5.2, post-training gains) and multimodal GLM-5.3-Flash on Aug 31. Source: https://docs.bigmodel.cn/cn/guide/models/text/glm-5.3

What to watch out for: GLM-5.3's headline +50% is an internal non-public benchmark, its per-MTok API price is not on the model page, thinking can no longer be disabled (a breaking cost change for simple calls), and exploit-generation capability doubled with no published usage policy. The "OX Alpha" stealth model fingerprinted to Zhipu is unconfirmed. Source: https://docs.bigmodel.cn/cn/guide/models/text/glm-5.3

Bottom line: GLM-5.3 beats Opus 4.8 on Zhipu's internal bench with ~42% of the tokens and posts the best CyberGym score recorded (84.5%, above Mythos 5), making Zhipu the open-weight token-efficiency and defensive-security leader; enterprise controls remain absent, so Western production adoption flows through third-party hosts. Source: https://docs.bigmodel.cn/cn/guide/models/text/glm-5.3

Key Terms

  • GLM-5.3 - Zhipu's new flagship (launched August 17): same base model as GLM-5.2, all gains from post-training. Text-only, 1M context, 128K max output, always-thinking with low/high/max effort (default max, disabling thinking no longer supported - a breaking API change). Source: Bigmodel – Glm 5.3
  • GLM-5.3-Flash - first native multimodal GLM-5 model (August 31): 320B total / 18B active parameters, sparse + linear attention hybrid, 1M context, 128K output, native image/video/file input. Source: Aliyun – Newly Released Models
  • Z.ai Code Bench - Zhipu's internal, non-public coding benchmark that places agents in complex local dev environments and measures end-to-end task completion; introduced with GLM-5.3 to avoid public-set contamination. Source: Bigmodel – Glm 5.3
  • Credit-based Coding Plan - the GLM Coding Plan's token-transparent billing (per-model input/cached/output coefficients, exact 5-hour and weekly credit limits); off-peak hours including all-day weekends deduct 50% credits. GLM-5.3 is fully rolled out to the plan. Source: Bigmodel – Overview
  • Vulnerability disclosure ledger - Zhipu's public ledger tracking vulnerabilities found by its models in real codebases, including affected projects, severity, CVE numbers, and age of the bug. Source: Bigmodel – Glm 5.3
  • "OX Alpha" (unconfirmed) - an anonymous stealth model on OpenRouter (August 20) scoring 80% DeepSWE Pass@1 with 1M context and native multimodality; independent fingerprinting (Ben Davis, reported by Local AI Zone) estimates 99% likelihood it is Zhipu's unreleased GLM-5.x multimodal flagship (~744B total / ~40B active). Zhipu has not confirmed attribution. Source: Local-Ai-Zone – Ai Updates August 2026

Latest Changes

Changes since the 2026-07 report.

July watch items verified:

  • GLM-5.3 launch: confirmed (was still-pending in July). GLM-5.3 launched August 17 with a technical report, full Coding Plan rollout, and SOTA claims on open-weight models for Terminal-Bench 3.0 and Agents' Last Exam. Sources: Bigmodel – Glm 5.3 , Z – Glm 5.3
  • Reward-hacking guard / cyber capability disclosures: partial. No new guard evaluations, but GLM-5.3's launch material discloses extensive new cyber-capability data (below) and a coordinated disclosure program for found vulnerabilities. Source: Bigmodel – Glm 5.3
  • Off-peak 50% credit deduction: confirmed continuing. The Coding Plan keeps 50% credit deduction in off-peak hours including weekends. Source: Bigmodel – Glm 5.3

New August changes:

  • New flagship (Aug 17): GLM-5.3, same base as GLM-5.2 with post-training gains: Z.ai Code Bench +50% vs GLM-5.2; Terminal-Bench 3.0 4.6 to 28.3; DeepSWE v1.1 46.2 to 66.9; Agents' Last Exam 23.8 to 28.5. Source: Bigmodel – Glm 5.3
  • Emergent cyber capability (Aug 17): CyberGym 84.5% (best recorded, above Mythos 5 at 83.8% and GPT-5.6 Sol at 83.6%); ExploitBench 54.4% vs GLM-5.2's 24.4% (Mythos 5: 78.0%); ExploitGym 105 exploits in 2 hours (Mythos 5: 181). In real-world testing with Chinese security teams: 2,436 vulnerabilities found in 269 projects, 1,097 medium-or-high severity, the oldest latent for ~40 years, now tracked in a public disclosure ledger. Source: Bigmodel – Glm 5.3
  • New model (Aug 31): GLM-5.3-Flash, first multimodal GLM-5 (320B/18B active), listed on Alibaba Bailian the same day; the z.ai homepage now brands itself "powered by GLM-5.3-Flash." Sources: Aliyun – Newly Released Models , Z
  • Breaking API change: GLM-5.3 cannot disable thinking; thinking.type: "disabled" requests fail - migrate to enabled with reasoning_effort: low. Source: Bigmodel – Glm 5.3
  • Protocol support: GLM-5.3 endpoints span OpenAI Chat Completions, OpenAI Responses, and Anthropic Message protocols (Claude Code-compatible); Coding Plan subscribers are temporarily limited to the OpenAI Chat Completions protocol. Source: Bigmodel – Glm 5.3
  • Marketplace distribution: GLM-5.3 and GLM-5.3-Flash added to Alibaba Bailian (Aug 17 / Aug 31). Source: Aliyun – Newly Released Models
  • Unconfirmed stealth model: "OX Alpha" on OpenRouter (Aug 20-27 free preview), fingerprinted as a probable unreleased Zhipu multimodal flagship. Not officially attributed. Source: Local-Ai-Zone – Ai Updates August 2026

Plans

Plan Price Notes
GLM Coding Plan Lite $16.2/mo (reg. $18) Credit-based; GLM-5.3 included; off-peak 50% deduction
GLM Coding Plan Pro $64.8/mo (reg. $72) Higher credit limits
GLM Coding Plan Max $144/mo (reg. $160) Highest limits
Team Edition Custom Central member/budget/permission management

USD prices carried over from July's ZCode listing; CNY domestic pricing separately maintained. Source: Zcode , Bigmodel – Overview

API Pricing

GLM-5.3 per-token API rates are published on the bigmodel.cn pricing pages with per-model credit coefficients in the Coding Plan docs; GLM-5.2's July rates (and the ~$1.40/$4.40 per MTok reference price seen on third-party platforms) carried over without a listed increase for 5.3 on the docs page fetched. Exact GLM-5.3 per-MTok rates: see Bigmodel – Overview for credit coefficients and Zhipu AI pricing for token rates.

Model Performance / Benchmarks

Benchmark GLM-5.3 GLM-5.2 Reference Notes
Z.ai Code Bench (max effort) 34.5% @ ~75K out-tokens/task 23.4% @ ~96K Fable 5: 39.5% Internal, non-public. Source: Bigmodel – Glm 5.3
Z.ai Code Bench (high effort) 31.4% @ ~50K - Opus 4.8: 29.5% @ ~120K Beats Opus 4.8 with ~42% of the tokens. Source: Bigmodel – Glm 5.3
Terminal-Bench 3.0 28.3 4.6 - Source: Bigmodel – Glm 5.3
DeepSWE v1.1 66.9 46.2 - Source: Bigmodel – Glm 5.3
Agents' Last Exam (CLI) 28.5 23.8 - Source: Bigmodel – Glm 5.3
CyberGym 84.5% 77.2% Mythos 5: 83.8%; GPT-5.6 Sol: 83.6% Best recorded score. Source: Bigmodel – Glm 5.3
ExploitBench 54.4% 24.4% Mythos 5: 78.0%; GPT-5.6 Sol: 76.5% Source: Bigmodel – Glm 5.3
ExploitGym (2h / 6h) 105 / 130 tasks 29 / 39 Mythos 5: 181 / 247 Source: Bigmodel – Glm 5.3

Latest News

GLM-5.3: Post-Training Leap (August 17, 2026)

GLM-5.3 keeps GLM-5.2's base weights and gains entirely from scaled post-training, including end-to-end synthesized task environments with verified binary reward signals. The efficiency story is the headline: at high effort it beats Claude Opus 4.8's internal-bench accuracy while emitting ~42% of the output tokens, and it closes most of the open-weight gap on terminal and deep-SWE benchmarks. Source: Bigmodel – Glm 5.3

Security Capability Emergence and Disclosure Program (August 2026)

Zhipu reports that vulnerability-exploitation capability grew faster than anticipated during post-training, documents the gap to Mythos 5 honestly, and has moved to a public vulnerability-disclosure ledger covering 2,436 real findings. Source: Bigmodel – Glm 5.3

OX Alpha: Stealth Preview Fingerprinted to Zhipu (August 20, 2026)

An anonymous model topping DeepSWE at 80% Pass@1 ran free on OpenRouter for a week; technical fingerprinting (tokenizer alignment with GLM-5.3, video-encoder token patterns matching GLM-5V-Turbo) points to an unreleased Zhipu multimodal flagship. Unconfirmed by Zhipu. Source: Local-Ai-Zone – Ai Updates August 2026

Community Signals

Aggregator coverage of GLM-5.3's launch framed it as the open-weight model closest to the cyber frontier and a token-efficiency leader; the OX Alpha mystery drew significant independent analysis. Source: Local-Ai-Zone – Ai Updates August 2026 . July's community benchmarks (Colibri on 32GB RAM, Martin Alderson's margin-collapse analysis, VAT benchmark) remain the standing third-party evidence for GLM-5.2 economics; no comparable third-party GLM-5.3 test has published yet. Sources: News – Item , Martinalderson – The Upcoming Ai Margin Collapse Part 1 Glm 5 2

Enterprise Readiness

Feature Available? Details
SSO (SAML) No Not documented (Team Edition has central management only).
SCIM No Not documented.
Audit logs No Not documented.
IP indemnity No Not documented.
Data residency No Not documented; international endpoint at open.bigmodel.cn.
HIPAA No Not advertised.
Air-gapped / on-prem Partial GLM-5.2 open weights enable self-hosting; GLM-5.3 weights not confirmed released.
SLA No Not published.
Admin controls (RBAC) Partial Team Edition member/budget/permission management. Source: Bigmodel – Overview
Anthropic-protocol compatibility Yes Native Anthropic Message endpoint enables Claude Code/Codex integration. Source: Bigmodel – Glm 5.3

Transparency Gaps

Gap Details Severity
GLM-5.3 API price Per-MTok rates for glm-5.3 are not on the model docs page; only Coding Plan credit coefficients are public. High
Z.ai Code Bench The headline +50% and Opus-beating claims rest on a non-public internal benchmark. High
OX Alpha attribution If Zhipu's, the stealth-release strategy bypasses benchmark accountability entirely; attribution remains unconfirmed. Medium
Exploit capability governance GLM-5.3 doubles exploit-bench scores with no published usage policy for offensive-security work (unlike Anthropic's CVP and OpenAI's Daybreak). Medium
Thinking always-on The removal of non-thinking mode raises cost per simple query; no discount tier is documented for low-effort calls. Low