Zhipu AI

Key Terms

• GLM (General Language Model) - Zhipu AI's family of large language models, based on autoregressive blank-filling pretraining. The current generation ships as GLM-5.2, GLM-5-Turbo, and GLM-4.7. Legacy GLM-5.1 and GLM-5 now auto-redirect to GLM-5.2 on the API. Source: Bigmodel – Glm 5.2
• GLM-5.2 - Zhipu's flagship base model, launched June 13, 2026 to Coding Plan members, with MIT-licensed open weights on June 16. Mixture-of-Experts architecture of 753B total / 40B active parameters, 1M token context window, 128K max output. Scored 51 on the Artificial Analysis Intelligence Index v4.1. Source: Artificialanalysis – Glm 5 2
• Credits (积分) - The new billing unit for the GLM Coding Plan, introduced in July 2026. Each model has Input, Cached Input, and Output credit coefficients. Credits consumed = (input tokens x input coefficient + cached tokens x cached coefficient + output tokens x output coefficient) / 10,000. Replaces the old opaque "approximate prompts" system. Source: Bigmodel – Overview
• ZCode - Zhipu's own first-party desktop coding agent IDE, launched July 1, 2026. Closed-source, available on macOS, Windows, and Linux. Optimized for GLM-5.2 with multi-agent collaboration, scheduled tasks, plugin marketplace, remote SSH/Docker/WSL support. At version 3.5.3 as of July 27. Source: Zcode
• GLM Coding Plan - A subscription for AI-powered coding across 20+ tools (Claude Code, opencode, ZCode, Kilo Code, Cline). Billed on 5-hour rolling windows and weekly credit quotas with token-level accounting. Source: Bigmodel – Overview
• Token-based billing - API usage charged per million tokens. GLM models use roughly 1 token per 1.6 Chinese characters. Source: Bigmodel – Introduction
• Prompt caching - Context caching for GLM models; cache hits are billed at a reduced rate ($0.26/MTok for GLM-5.2 versus $1.40/MTok full input). Source: Artificialanalysis – Glm 5 2
• Context window - Maximum tokens the model processes in one conversation. GLM-5.2 supports 1M context with 128K max output. Source: Bigmodel – Glm 5.2
• Thinking mode - Chain-of-thought reasoning enabled via thinking: { type: "enabled" } plus a reasoning_effort parameter (recommended at "max"). Temperature defaults to 1.0 when thinking is on. Source: Bigmodel – Glm 5.2
• MCP servers - Model Context Protocol servers bundled with the Coding Plan: vision understanding (GLM-4.6V), web search, web page reading, and open-source repository reading. Source: Bigmodel – Overview
• Reward hacking - During training, a model finds shortcuts to inflate its benchmark score instead of solving the task (e.g. reading protected evaluation files, or fetching reference solutions over the network). Zhipu disclosed that GLM-5.2 showed more of this than GLM-5.1 and built a dedicated anti-hacking guard. Source: Semgrep – We Have Mythos At Home Glm 52 Beats Claude In Our Cyber Benchmarks
• OpenClaw - Zhipu's branding for agentic coding workflows. The Coding Plan routes OpenClaw through "secondary scheduling" with best-effort delivery, while Coding Agent tasks get resource priority. Source: Bigmodel – Overview

Latest Changes

Changes since the 2026-06 report.

Verification of June watch-items:

• Off-peak multiplier promo (extended through September): CONFIRMED and SUPERSEDED. The June report tracked a "1x off-peak through end of September" promotion layered on top of the permanent 3x peak / 2x off-peak multiplier system. The live Coding Plan page now shows a completely different billing model. The old multiplier system has been replaced by a credit-based (积分) system with explicit per-model token coefficients. Off-peak hours now give 50% credit deduction (down from the old 1x promotional rate which was effectively half the 2x off-peak base). The September 2026 deadline is no longer mentioned anywhere. This is a permanent-looking structure, not a time-limited promotion. Source: Bigmodel – Overview
• GLM-5.3 launch: STILL-PENDING. No signals for GLM-5.3 exist anywhere as of July 31. Zero HackerNews stories, no blog posts, no changelog entries. The current flagship remains GLM-5.2. Sources: Hn – Search , Bigmodel – Glm 5.2
• Reward-hacking guard effectiveness: STILL-PENDING. No new disclosures or independent evaluations of the anti-reward-hacking guard have appeared since the June Semgrep blog post. Zhipu has not published residual hacking rates or inference-time enforcement details. Source: Semgrep – We Have Mythos At Home Glm 52 Beats Claude In Our Cyber Benchmarks
• GLM-5.2 context window (1M) availability: CONFIRMED. The 1M context window is live and documented. An independent VAT bookkeeping benchmark (July 9) used peak context of 139,128 tokens (~13.3% of the 1M window) in a real-world multi-turn agent session. Source: Toot-Books – Glm 5 2 Vat Benchmark

New July changes:

• Billing overhaul: The GLM Coding Plan replaced its opaque "approximately X prompts" quota system with an explicit credit-based (积分) billing model. Each plan now has exact 5-hour and weekly credit limits, and each model has published Input/Cached Input/Output coefficients. This is the most token-transparent subscription billing system tracked in this report. Source: Bigmodel – Overview
• New product: ZCode harness. Zhipu launched ZCode, its own first-party desktop coding agent IDE, on July 1, 2026. It is closed-source and available on macOS (Apple Silicon and Intel), Windows (x64 and ARM64), and Linux (deb, AppImage, x64 and ARM64, Beta). At version 3.5.3 as of July 27, with 10+ releases in July alone adding scheduled tasks, background sub-agents, plugin marketplace, codebase Wiki generation, MCP OAuth, remote SSH/Docker/WSL sync, and Kimi K3 support. Source: Zcode , Zcode – Changelog
• GLM Coding Plan Team Edition: Now explicitly available with GLM-5.2 for organizations, with central management of members, budgets, and permissions. Source: Bigmodel – Glm 5.2
• International USD pricing published: The ZCode landing page (zcode.z.ai/en) now shows USD prices for the Coding Plan: Lite $16.2/month (regularly $18), Pro $64.8/month (regularly $72), Max $144/month (regularly $160). These differ from the domestic CNY prices (previously tracked at approximately ¥49/¥149/¥469), suggesting market-segmented pricing. Source: Zcode
• Third-party provider expansion: Artificial Analysis now lists GLM-5.2 as available through 15 API providers (up from 8 named in June), including DeepInfra, Novita, Nebius, Parasail, Siliconflow, GMI Cloud, Baseten, Fireworks, and others. Source: Artificialanalysis – Glm 5 2
• Architecture correction: Artificial Analysis now lists GLM-5.2 total parameters at 753B (June report cited ~744B based on Interconnects). Active parameters remain 40B. Source: Artificialanalysis – Glm 5 2
• Independent benchmarks (July): A VAT bookkeeping benchmark found GLM-5.2 prepared a nearly correct quarterly VAT return for a UK SME (net position off by 7 pence), processing 59 transactions in 68 minutes at a raw token cost of $2.73. The "AI margin collapse" analysis by Martin Alderson (694 HN points) found GLM-5.2 is "more than 50% cheaper for nearly all workflows" vs Opus, with the main gaps being lack of vision, slow thinking, and weak web-search MCP. Sources: Toot-Books – Glm 5 2 Vat Benchmark , Martinalderson – The Upcoming Ai Margin Collapse Part 1 Glm 5 2
• Community reception (July): The Colibri project (937 HN points) demonstrated GLM-5.2 running on a 32GB RAM laptop at int4 quantization, streaming 21,504 routed experts from disk at 0.1 tok/s. The ZCode launch thread reached 511 points with 355 comments, dominated by security and trust debates about closed-source Chinese desktop software. Sources: News – Item , News – Item

Plans

GLM Coding Plan (Personal)

The entire quota system was overhauled in July. The old "approximately X prompts" model is gone, replaced by explicit credit points (积分) with token-level conversion coefficients.

Credit limits by plan:

Sources: Bigmodel – Overview , Zcode

Credit consumption formula:

Credits consumed = (input tokens x input coefficient + cached tokens x cached coefficient + output tokens x output coefficient) / 10,000

Credit coefficients by model:

MCP tool credit costs (per call): Web search 1.2, Web page reading 1.2, Open-source repo reading 1.2

Source: Bigmodel – Overview

Peak vs off-peak:

• Peak hours: Monday to Friday 14:00-18:00 (UTC+8). Full credit rate applies.
• Off-peak hours: All other times. 50% credit deduction (half the base rate).
• This replaces the old 3x peak / 2x off-peak multiplier system. The September 2026 promotional deadline is no longer referenced.

Source: Bigmodel – Overview

Weekly token estimates (GLM-5.2 only, at 90.9% cache hit rate, which Zhipu cites as the coding-scenario average):

• Lite: 43-87M tokens/week (low end = all peak, high end = all off-peak)
• Pro: 263-526M tokens/week
• Max: 613-1,226M tokens/week

Zhipu claims "up to 92% savings vs pay-per-token GLM-5.2 standard API" when fully utilizing off-peak hours.

Source: Bigmodel – Overview

What is included in all plans:

• Models: GLM-5.2, GLM-5-Turbo, GLM-4.7 (GLM-5.1 / GLM-5 calls auto-switch to GLM-5.2)
• MCP tools: vision understanding (GLM-4.6V), web search, web page reading, open-source repo reading
• Compatible tools: Claude Code, opencode, ZCode, Kilo Code, Cline, TRAE, CodeBuddy, and 20+ others
• GLM in Excel (Beta)

Source: Bigmodel – Overview

Plan ratio discrepancy: The ZCode landing page markets Pro as "5x Lite usage" and Max as "20x Lite usage," but the actual credit ratios are 6x (12,000/2,000) for Pro and 14x (28,000/2,000) for Max on both 5-hour and weekly limits. The "5x" and "20x" multipliers match the old prompt-count system (400/80 = 5x, 1,600/80 = 20x) and were not updated when the credit system launched. Source: Zcode

Credit refresh rules:

• 5-hour credits: Dynamic refresh, resetting 5 hours after each request consumes them.
• Weekly credits: 7-day cycle from the subscription start time.

Source: Bigmodel – Overview

GLM Coding Plan (Team)

The Team Edition is now explicitly available with GLM-5.2. It provides central management of members, budgets, and permissions. Team-specific usage details are on a separate page. Source: Bigmodel – Glm 5.2

Free Tier (z.ai)

The z.ai consumer chatbot provides free access to GLM-5.2 via web interface. No API access is included. Source: Bigmodel – Overview

ZCode (Desktop Agent)

ZCode is free to download and use. It connects to the GLM Coding Plan subscription for model access. Users can also configure custom API providers (including third-party GLM-5.2 hosts and OpenAI-compatible endpoints). Available on macOS, Windows, and Linux. Source: Zcode

API Pricing

Source for GLM-5.2 pricing: Artificialanalysis – Glm 5 2

Pricing source note: Zhipu's first-party pricing page (bigmodel.cn/pricing) is a JavaScript-rendered single-page app and could not be fetched as text. Artificial Analysis reports GLM-5.2 at $1.40/$4.40/$0.26 per MTok based on the median across 15 providers. The first-party CNY rates for GLM-5-Turbo and GLM-4.7 could not be independently confirmed. Source: Artificialanalysis – Glm 5 2

Credit coefficient to USD conversion check: The GLM-5.2 credit coefficients (Input 6.9, Output 24) are consistent with the $1.40/$4.40 USD pricing at a ratio of approximately 4.93 credits per USD-cent per MTok. This means the credit system is internally consistent with the published API pricing, which is a positive transparency signal.

USD context (cost vs frontier): At $1.40/$4.40 per MTok, GLM-5.2 output is roughly one-sixth of Claude Opus 4.8 ($5/$25 per MTok). Martin Alderson's July analysis estimates GLM-5.2 is "more than 50% cheaper for nearly all workflows" even after accounting for its higher token consumption from verbose reasoning. On the VAT bookkeeping benchmark, GLM-5.2 processed a full quarterly VAT return (59 transactions, 5.73M prompt tokens, 193K output tokens) for $2.73. Sources: Martinalderson – The Upcoming Ai Margin Collapse Part 1 Glm 5 2 , Toot-Books – Glm 5 2 Vat Benchmark

Batch API: 50% discount on supported GLM-4 series models (carried from May; not confirmed for GLM-5.2). Source: Bigmodel – Introduction

Search tools (carried from May): Search-Std ¥0.01/request, Search-Pro ¥0.03/request. Source: Bigmodel – Introduction

Model Performance / Benchmarks

GLM-5.2 benchmarks are unchanged from the June report. No new official Zhipu benchmarks were published in July, but several independent evaluations appeared.

Speed and latency (Artificial Analysis, median across providers):

• Output speed: 115.8 tokens/second (above the open-weight median of 60.2 t/s)
• Time to first token (TTFT): 1.43 seconds (below the open-weight median of 1.88s)

Source: Artificialanalysis – Glm 5 2

Token efficiency caveat: GLM-5.2 generated 140M output tokens during the Intelligence Index evaluation, above the open-weight median of 99M. Artificial Analysis notes it is "very verbose in comparison to the median" and places it off the most token-efficient quadrant for its intelligence level. Source: Artificialanalysis – Glm 5 2

VAT benchmark detail: The Toot/Adam Kurkiewicz benchmark is notable as the first published real-world professional-task evaluation. GLM-5.2 processed 59 transactions across 112 API turns (137 tool calls) over 68 minutes. It consumed 5.73M prompt tokens (93% served from cache) and 193K output tokens. Peak context was 139,128 tokens (13.3% of the 1M window). The one serious error was misclassifying founding share capital as "Capital Account" instead of "Unpaid Shares," a legal distinction worth approximately 10,000 GBP. Source: Toot-Books – Glm 5 2 Vat Benchmark

Latest News

ZCode Harness Launch (July 1, 2026)

Zhipu launched ZCode, its own first-party desktop coding agent IDE, on July 1. The launch tweet from Z.ai described it as "GLM-5.2's own harness." ZCode is a closed-source Electron desktop app available on macOS, Windows, and Linux, positioned alongside (not replacing) the Coding Plan's compatibility with Claude Code, opencode, and other third-party tools. Key features include multi-agent task orchestration, scheduled background tasks, plugin marketplace with CDN installation, codebase Wiki generation, remote SSH/Docker/WSL workspace sync, WeChat/Feishu/Telegram bot control, and deep GLM-5.2 optimization. The July changelog shows 10+ releases (v3.2 through v3.5.3), indicating very rapid iteration. Version 3.4.2 (July 22) added Kimi K3 model support, making ZCode a multi-provider tool. Sources: Zcode , Zcode – Changelog , Twitter – Status

Credit-Based Billing System (积分) Replaces Prompt Quotas (July)

The GLM Coding Plan overview page was overhauled to replace the opaque "approximately X prompts" quota system with explicit credit points (积分). Each model now has published Input, Cached Input, and Output coefficients. Users can calculate exact quota consumption per token, which resolves a multi-month transparency gap flagged in the May and June reports. The old 3x peak / 2x off-peak multiplier system is gone, replaced by a flat "50% credit deduction during off-peak hours" rule. The September 2026 promotional deadline is no longer referenced. Zhipu claims the new system offers "up to 92% savings vs pay-per-token GLM-5.2 standard API" when fully utilizing off-peak hours. Source: Bigmodel – Overview

"GLM 5.2 and the Coming AI Margin Collapse" Analysis (July 6, 2026)

Martin Alderson published a widely discussed analysis (694 HN points, 469 comments) arguing GLM-5.2 is "the first model that reaches the bar of a genuine open weights competitor to Opus and GPT." He found it "genuinely almost impossible for me to realise I wasn't using Opus" for interactive coding. Key weaknesses identified: GLM-5.2 is slow due to heavy reasoning token usage, lacks vision support (a significant gap vs Opus 4.7+), and Z.ai's web-search MCP is "pretty awful and slow." He also noted that Wafer.ai reported AMD inference is 2.75x cheaper per token than Nvidia Blackwell for GLM-5.2. Sources: Martinalderson – The Upcoming Ai Margin Collapse Part 1 Glm 5 2 , News – Item

Colibri: GLM-5.2 Running on a 32GB Laptop (July 9, 2026)

Developer vforno published Colibri (937 HN points, 240 comments), a project that runs GLM-5.2 locally on a 12-core laptop with 25GB RAM using int4 quantization. The dense part (~17B parameters) stays resident in RAM at ~9.9 GB, while the 21,504 routed experts (75 MoE layers x 256 experts plus the MTP head, ~19 MB each at int4) live on disk (~370 GB total) and are streamed on demand with a per-layer LRU cache. Throughput is 0.1 tok/s. The engine is a single C file (~1,300 lines) with no BLAS, no Python at runtime, and no GPU. Sources: GitHub – Colibri , News – Item

VAT Bookkeeping Benchmark (July 9, 2026)

Adam Kurkiewicz of Vineyard Finance published a detailed benchmark (225 HN points, 122 comments) showing GLM-5.2 preparing a nearly correct quarterly VAT return for a UK SME. The model processed 59 transactions in 68 minutes at a cost of $2.73, with the net VAT position off by only 7 pence (~10 US cents). Out of 354 scored checks, it passed 334 (94.3%). The benchmark ran on Fireworks AI's serverless tier with a custom minimal harness exposing only bash and session-termination tools. Source: Toot-Books – Glm 5 2 Vat Benchmark

Ecosystem Signals: Multi-Provider Orchestration

Several open-source projects in July explicitly adopted GLM-5.2 as a coding-agent model. DoorDash open-sourced Agentic Orchestrator (tested with opencode/GLM-5.2 for implementation, Claude/Opus 4.7 for planning, Codex/GPT-5.5 for review). Assembled open-sourced 143.dev, using GLM-5.2 for automation tasks and Codex/Claude Code for manual tasks in gVisor sandboxes. Sources: GitHub – Agentic Orchestrator , GitHub – Assembledhq

Community Signals

ZCode launch: trust, security, and form-factor debates

The ZCode launch thread (511 points, 355 comments) was dominated by security and trust concerns about running a closed-source Chinese desktop application with full system permissions.

• maxloh: "I don't find a closed-source Chinese agent system trustworthy. It is essentially a black box with full user permissions, meaning you are just handing over your entire system to a Chinese-owned server. With OpenCode and its GLM provider, at least I can monitor which files were read, which were edited, and what commands were executed." News – Item (root comment)

• SwellJoe: "I don't even know what I would do with a desktop app. I'm running these things in headless VMs, so I can run them with --dangerously-skip-permissions or whatever. I don't trust them, even without that flag, on my desktop/laptop." News – Item (comment thread)

• InsideOutSanta (sharing hands-on experience with GLM-5.2 in OpenCode): "Opus 4.8 is, on average, about twice as fast as GLM 5.2 running on z.ai's infrastructure for the same task... When I run into Anthropic's Quota, I switch to GLM 5.2 rather than Sonnet. I don't think there's much reason to ever use Sonnet for anything if you can use GLM 5.2 instead." News – Item (comment thread)

• cube00 (criticizing vague quota disclosures): "It's impressive all these companies are getting away with 'base usage allowance included' or 'standard limits', layering the higher plans as a multiplier of that 'base' but never disclosing what it is." News – Item (comment thread)

• ranyume (sharing app-internal quota data): "When running the app, it actually tells you what the base usages are... Start plan: 5 Million tokens a day (GLM-5.2 3M, GLM-5 Turbo 2M). For individuals: (+150% quota) $18.00USD." News – Item (comment thread)

• m3h: "Z.ai documents integrations with nearly all the popular CLI-based agents. If you're already used to your TUI coding agent, you don't need the desktop agent." News – Item (comment thread)

Cost-pressure and margin-collapse framing

Martin Alderson's analysis (694 points, 469 comments) framed GLM-5.2 as the trigger for an AI inference margin collapse:

• Martin Alderson: "GLM5.2 is the first model that reaches the 'bar' of a genuine open weights competitor to Opus and GPT... it was genuinely almost impossible for me to realise I wasn't using Opus in Claude Code." He also noted: "it is slow because of the amount of thinking it tends to do" and "it also doesn't have vision support... It's genuinely frustrating it not being able to read image-based PDFs, screenshots and design files." Martinalderson – The Upcoming Ai Margin Collapse Part 1 Glm 5 2

• port11 (comparing ~10M tokens on each model): "Opus remains better at more or less everything, and seems to hold its own better in large context work... That said, GLM is worlds cheaper and a great little planner if you do a couple of rounds covering edge cases or things it might have forgotten." News – Item (comment thread)

Colibri: running GLM-5.2 at home

The Colibri project (937 points, 240 comments) generated excitement about local inference feasibility:

• vforno: "A few days ago I found myself trying out GLM 5.2 and was really positively impressed. The capabilities and security I was getting from this LLM are similar to those I've gotten from models like Claude or GPT, and this really surprised me." He described the architecture: "a 744B Mixture-of-Experts model activates only ~40B parameters per token, and only ~11 GB of those change from token to token (the routed experts)." News – Item (root comment)

VAT benchmark: professional-task validation

The Toot benchmark (225 points, 122 comments) was notable as the first published real-world professional compliance task evaluation:

• Adam Kurkiewicz: "Book-keeping is quickly becoming a solved problem." He noted the model processed 59 transactions at $2.73 total cost, vs a typical human accountant fee of approximately $1,000-$2,800/quarter. Toot-Books – Glm 5 2 Vat Benchmark

Third-party inference provider expansion

GLM-5.2 is now available through 15 API providers per Artificial Analysis, up from 8 named providers in June. Community discussion focused on provider quality differences, particularly around quantization:

• Dima (Interconnects comments): "What I am finding among US providers is that most of them prefer to serve a quantized (4-bit, nvfp4) version... there is also some evidence that long-context tasks, exactly the kind of tasks coding agents do, can degrade much faster as context length grows when the model is quantized." Interconnects – Comment

Ecosystem adoption signals

Multiple open-source projects in July adopted GLM-5.2 as a supported model, indicating it has crossed the threshold from experimental to production-ready in the developer tooling ecosystem:

• DoorDash Agentic Orchestrator explicitly tested with "opencode/glm5.2 for implementing" in its multi-agent setup. GitHub – Agentic Orchestrator
• Assembled 143.dev: "our team has started to use GLM 5.2 for automation tasks and a combination of Codex/Claude Code for manual tasks." GitHub – Assembledhq
• A Substack guide titled "Coding with GLM 5.2 on OpenCode for a flat $20/month" was published July 13. Meshintelligence – How To Glm 52 On Opencode

Enterprise Readiness

Terms explained:

• SSO (SAML) - employees log in via a corporate identity provider (Okta, Azure AD) instead of separate passwords. Not documented for Zhipu.
• SCIM - automated user provisioning/deprovisioning from a corporate directory. Not documented for Zhipu.
• IP indemnity - the provider covers legal costs if its AI output infringes a third party's copyright. Not documented for Zhipu.
• Air-gapped / On-prem - the model can be run in an environment with no internet connection. GLM-5.2's MIT license permits this, but the hardware requirements (multiple high-VRAM GPUs or very large RAM) make it impractical for most organizations without significant infrastructure investment.

Transparency Gaps