On file for Z.ai (GLM)
- Paid · GLM Coding Plan from ~$3/month; API usage billed per token
- Teams running agentic coding or bulk generation who need frontier-adjacent quality at a fraction of US per-token pricing
- AI Agents & Workflow Builders
- Written review
What is Z.ai (GLM)?
Z.ai is the developer platform for Zhipu AI's GLM model family. The current lineup spans GLM-5.3 (flagship), GLM-5.3-Flash and GLM-5.3-FlashX (fast, low-cost variants), and the previous-generation GLM-4.5, which remains available with open weights. All models are served through a unified API gateway documented at docs.z.ai, with a 128k context window across the 5.3 family, structured function calling, and an optional deep-reasoning mode that emits explicit reasoning tokens before the final answer.
Architecturally, the 5.3 family runs on sparse Mixture-of-Experts serving with a hybrid reasoning design: you toggle thinking per request, and the model either answers directly or burns reasoning tokens first. The serving stack exposes three interfaces: an OpenAI-compatible chat completions endpoint, a native Anthropic-compatible /v1/messages endpoint, and a web console for key management.
The Anthropic-compatible endpoint is the reason most operators care. It means Claude Code and Cursor point at Z.ai with a base-URL override and no proxy shim — the tool-call schema, streaming format, and stop-reason semantics are translated natively on Z.ai's side. The problem it solves is specific: frontier agentic coding burns tokens aggressively, and Western frontier pricing makes retry-heavy agent loops expensive. GLM-5.3-Flash lists at roughly $0.0001 per 1k tokens — one to two orders of magnitude below Claude or OpenAI output pricing.
Who is Z.ai (GLM) for?
Benefits most:
- Engineering teams running Claude Code or Cursor daily who hit subscription or token ceilings and need more agent throughput per dollar
- High-volume pipeline workloads — classification, extraction, summarization, log triage — where GLM-5.3-Flash's price dominates quality deltas
- Bootstrapped startups and agencies prototyping agents, where retry-heavy loops make frontier-priced models cost-prohibitive
- Teams maintaining internal coding assistants with flat, predictable monthly budgets ($18/month tiers well against per-token surprise bills)
Should skip:
- Regulated enterprises (US/EU healthcare, fintech, public sector) that require SOC 2 Type II reports and contractual data residency — Z.ai's compliance documentation is thin
- Teams that need >128k context for whole-monorepo agent sessions; Claude-class 200k+ windows still matter here
- Workloads with low retry tolerance on long-horizon refactors, where Claude Sonnet's tool-call reliability still justifies its price
- Anyone who needs open weights for the current flagship; only GLM-4.5 is self-hostable, the 5.3 family is API-only
Core Capabilities and Workflows
Model lineup
| Model | Role | |---|---| | GLM-5.3 | Flagship. Hybrid reasoning, strongest agentic/tool-calling performance in the family | | GLM-5.3-FlashX | Mid-tier. Low latency, modest reasoning budget, cheaper than flagship | | GLM-5.3-Flash | Volume tier. ~$0.0001/1k tokens blended. Fast, shallow reasoning by default | | GLM-4.5 | Legacy. Open weights (MIT) on Hugging Face; still capable for extraction and routing |
Anthropic-compatible gateway
The primary workflow for coding agents. No shim library required:
export ANTHROPIC_BASE_URL="https://api.z.ai/api/anthropic"
export ANTHROPIC_AUTH_TOKEN="your-zai-api-key"
claude
Claude Code's sub-agents, TodoWrite, and edit-tool flows work against the translated schema. Cursor works via the OpenAI-compatible base URL (https://api.z.ai/api/paas/v4/chat/completions) with a model-name override. Expect occasional edge cases in tool-call argument validation on the Flash tiers — see trade-offs below.
Function calling and structured output
Parallel tool calls and JSON-schema-constrained output are supported natively. Schema enforcement is stricter on GLM-5.3 than Flash; on Flash we observed malformed argument JSON roughly 2–3% of the time in long tool chains, which agentic frameworks handle via retry but which burns credits.
Reasoning tokens
Setting thinking: true emits reasoning tokens before the final response. These bill as output tokens on the API. Reasoning overhead on hard prompts runs 2–4x the visible output length — material on GLM-5.3, negligible on Flash pricing.
Prefix caching
Automatic. Repeated system prompts and stable context prefixes bill at roughly 20% of standard input rate on cache hits. On the Coding Plan, cached prefixes also reduce credit burn, which matters on monorepos where Claude Code re-sends large stable context every turn.
Pricing and Plans
Pay-as-you-go (API, prepaid wallet)
Observed list prices, verified 2026-11. Zhipu has historically adjusted list pricing quarterly — re-check before budgeting.
| Model | Input /1M | Output /1M | |---|---|---| | GLM-5.3 | $1.00 | $3.20 | | GLM-5.3-FlashX | $0.20 | $0.80 | | GLM-5.3-Flash | $0.06 | $0.22 | | GLM-4.5 | $0.60 | $2.20 |
- Cache-hit input bills at ~20% of list input rate.
- Minimum prepaid wallet top-up: $10.
- Reasoning tokens bill at the output rate.
Coding Plan — $18/month
- 10,000 credits per week, reset every Monday 00:00 UTC. No rollover.
- 1.5x usage bonus currently applied to all credit consumption — effectively 15,000 credits/week. Treat this as promotional; budget against the base 10,000.
- Credits are consumed per request, not per token. A standard Claude Code prompt on ~40k context ≈ 1 credit. Prompts on >100k context or with deep reasoning enabled ≈ 3–6 credits. Practical throughput: roughly 2,500 aggressive agent prompts/week.
- Prefix caching is included and reduces credit burn on repeated context.
- Concurrency is capped (~3 parallel agent sessions observed); peak-hour queuing occurs.
- No automatic overage billing. When credits exhaust, requests fail with a quota error unless you explicitly toggle fallback to the pay-as-you-go wallet.
Hidden costs
- FX fees: Billing is USD-only. Non-US operators eat a 1–3% issuer FX spread on every charge — the $18 plan lands closer to $18.30–$18.60 on EUR/GBP statements. Some issuers flag the merchant (a China-registered entity) and decline charges outright.
- VAT/GST is not included in listed prices where applicable.
- Weekly pool granularity: release-heavy weeks exhaust credits; quiet weeks waste them. There is no monthly pooling option.
- Reasoning-token inflation on the API: deep thinking mode can triple effective output cost on GLM-5.3.
Where does Z.ai (GLM) fall short?
- 128k context ceiling. Claude Code's auto-compact triggers noticeably earlier than on 200k-window models. Large monorepo sessions degrade into more frequent compaction cycles.
- Agentic reliability on complex chains. On multi-step refactors involving 15+ tool calls, GLM-5.3 required roughly 20–30% more retries than Claude 3.5 Sonnet in our runs. Flash drifts more. The per-token savings partially offset the retry cost, but wall-clock time suffers.
- Rate-limit gates for new accounts. Fresh API keys start at low RPM/TPM tiers and are upgraded based on cumulative spend or manual review. Expect a week of friction before production-grade concurrency.
- Tool-call schema edge cases on Flash. Occasional malformed JSON arguments in parallel tool calls. Fine for supervised use; a liability in unattended pipelines.
- Promotional dependency. The 1.5x credit bonus is time-limited. Budget models built on 15,000 weekly credits break when it lapses.
- Compliance posture. No public SOC 2 Type II, no signed DPA template at time of review, and inference regions spanning China plus international
Henryk Wexel runs Quota Engine. He spent a decade selling B2B software and building outbound systems — as a founder, as a seller at Silverfort and Cradlepoint, and as the person on the hook for pipeline numbers. He writes the reviews here because he has bought, deployed, argued about and cancelled most of the tools in this index. Based in Palma de Mallorca.